Intelligence Operating System
NEO AI x100
One assistant. An entire AI team.
Talk naturally. NEO selects the models, coordinates agents and checks the work. One conversation. All that power.
An Intelligence Operating System designed to make leading models, agents, tools, data, memory, verification and compute - including compute already owned by its users - operate as one coordinated intelligence system, and to multiply useful intelligence per unit of compute.
COREORCHESTRATION
- GPT
- Agents
- Claude
- Tools
- Gemini
- Memory
- Grok
- Verifiers
- Kimi
- Data
- Qwen
- Edge compute
- DeepSeek
- Private models
- Specialist engines
- Enterprise systems
The problem
AI is powerful. Using it is hard.
Too many subscriptions
Separate tools. Separate bills.
Too much prompt trial and error
You keep rewriting instead of working.
Too many models to choose from
You become the AI operator.
Your goal gets lost in the tools.
The solution
One AI assistant for all your needs
Tell NEO your goal. It asks, plans and works with you through completion.
YouHelp me prepare an investor pitch.
NEOWho is the audience? What should they remember?
Work you can useAn investor pitch deck. Story. Slides. Revisions. Reviewed with you.
No subscription juggling
No prompt training
No model guesswork
Supported models and usage depend on the NEO plan. Journey shown is illustrative.
What NEO is
Beyond the AI assistant
NEO AI is an Intelligence Operating System designed to go beyond traditional AI assistants such as ChatGPT or Claude. Instead of relying on a single model, NEO orchestrates multiple leading AI systems - GPT, Claude, Gemini, Grok, Kimi, Qwen, DeepSeek, private models and specialist engines - together with autonomous agents, tools, data sources, memory, verification, enterprise controls and heterogeneous compute - from the intelligence already present on the user's own device to private infrastructure and frontier clouds.
ORCHESTRATION
- ModelsGPT · Claude · Gemini · Grok
- ModelsKimi · Qwen · DeepSeek
- Private modelsenterprise · specialist engines
- Autonomous agentsresearch · build · critique
- Tools & data sourcesAPIs · databases · documents
- Memorypersistent · verified
- Verificationindependent verifiers
- Enterprise controlssecurity · policy · jurisdiction
Not simply more capable AI
The objective of NEO is not simply to make AI more capable, but to make the entire intelligence system dramatically more efficient, more reliable and more powerful.
A long-term architecture
This is the principle behind NEO x100: a long-term architecture designed to multiply effective intelligence while reducing the amount of expensive compute required to obtain it.
Rather than solving every problem by sending everything to the largest and most expensive frontier model, NEO decomposes a Mission into specialized units of work and dynamically decides which model, agent, tool, memory, data source and compute path should handle each one - and where it should execute.
Simple on the surface
Open NEO. Ask LILI. Receive the result.
A private user, a professional or an enterprise team opens NEO the way they would open any leading AI assistant - one conversation, one objective, one result. The Intelligence Operating System does the rest: orchestration, models, agents, routing, local and cloud compute allocation, memory and verification stay beneath the surface unless the user chooses to look.
- Open NEOone conversation
- Ask LILIone Mission
- Accept a profilenever a model list
- NEO does the workbeneath the surface
- See progressif wanted
- Receive the resultverified
The user never needs to understand orchestration, models, agents, routing, task graphs, APIs or infrastructure.
Under the surface
GPT - Claude - Gemini - Grok - Kimi - Qwen - DeepSeek - private
research - build - critique - verify
APIs - runtimes - connectors
persistent - verified - provenance
independent cross-check
capability - cost - policy
cascades - reuse - allocation
documents - sources - systems
What a private user sees
- A conversation with LILI and a visible recommendation
- Sources counted and claims verified - with the evidence one click away
- One decision card when a Human Gate is required
- An Intelligence Profile the user can accept or change - never a model list
- Progress, Output and Context in the inspector, if wanted
As immediately understandable as a consumer assistant
FOCUS keeps the conversation and the decision. Everything NEO can do remains available underneath; nothing has to be configured to start.
NORMAL and JARVIS reveal more
The same Mission can be observed with execution plan, agents, evidence and, in JARVIS, the full command core - without ever becoming a different application.
Not by unsupported benchmarks
NEO differentiates on how work is organized - orchestration, verification, provenance, authority - not on claimed superiority over any named product.
Explore NEO AI
One system, seen from every side
You do not open LILI. LILI is already there.
Live demoOne Mission state. Three ways to see it.
x100 TechnologyFrom one Mission to thousands of coordinated sub-tasks
ArchitectureCloud when needed. Local when sufficient. NEO decides.
SolutionsAn AI organization, not an AI assistant
Business modelSeveral layers, two businesses
The key strategic difference
Summarized very simply
A traditional assistant gives you a powerful AI assistant. NEO AI is designed to give you an entire AI organization - while continuously optimizing which models, agents, tools, memory and compute resources are used for every individual piece of work.
A powerful AI assistant.
The user interacts with a provider's assistant and model stack.
An entire AI organization.
Models, agents, tools, memory, verification and compute - user-owned, private, NEO-controlled, frontier - orchestrated for every piece of work, wherever it is best executed.
NEO x100 - the longer-term R&D objective
- ✗ not 100x more parameters
- ✗ not 100x more GPUs
- ✗ not brute-force intelligence
but the long-term objective of orders-of-magnitude improvement in verified intelligence-per-compute through orchestration, specialization, parallelism, verification, memory, reuse, edge execution, hardware-aware routing and adaptive resource allocation.
- Orchestration
- Specialization
- Parallelism
- Verification
- Memory
- Compute reuse
- Adaptive allocation
- = NEO x100
Compounding multipliers - not parameters, not GPUs.
More intelligence. Greater reliability. Faster execution - with substantially less unnecessary compute. Models are inputs. Compute is an input. The NEO system is the asset.
Other assistants answer when you ask. LILI is already there.
- Hands-free voiceSay "Hey LILI". Keep talking.
- Present 24 / 7Watches. Alerts. Acts for you.
- You stay in controlAlways-on never bypasses a Human Gate.
LILI Always-on
You do not open LILI. LILI is already there.
An assistant waits to be asked. LILI Always-on stays present: it listens for your voice, follows the Missions you launched, watches what you entrust to it, and comes to you the moment something needs you - on every device, day and night.
ALWAYS-ON
24 / 7
- iPhone · Androidon every device
- Listenshands-free voice · wake word "Hey LILI"
- Mac · Windowson every device
- Watchesyour Missions and what you entrust
- Webon every device
- Alertsthe moment something needs you
- Workspaceon every device
- Actsproactively · inside Human Gates
Say "Hey LILI". Keep talking.
- A wake word instead of an app to open
- A continuous spoken conversation - ask, interrupt, refine, decide
- On the phone, the computer and the other devices you authorize
- The same Mission whether you speak, type or switch device
Work that keeps moving while you do not
- Follows every running Mission to completion
- Watches the signals you entrust to it - a KPI, an inbox, a deadline, a build
- Alerts you the moment a decision is yours to take
- Acts proactively on what you have already authorized
A day with LILI Always-on
- 02:14overnight Mission finishes
- 07:30spoken morning brief
- 11:05approval asked · Human Gate
- 16:40KPI alert · you answer by voice
- 21:30tomorrow prepared
Always-on never bypasses a mandatory approval
On, paused or off - per device, at any time
Presence is a permission, never training consent
FOCUS · NORMAL · JARVIS show the same work
Other assistants answer when you ask. LILI is designed to be there before you do.
Three projections of one state
One Mission state. Three ways to see it.
FOCUS, NORMAL and JARVIS are three projections of one Mission state - not three applications. Changing mode changes what the user sees, never what NEO is doing: execution is not restarted, no second Mission is created and the backend truth does not change. Execution location is equally independent of projection; the Mission belongs to NEO, not to the device that started it.
MISSION
STATE
- Progress
- Output
- Context
- FOCUSminimal surface · the essential
- JARVIScommand-center depth
- NORMALthe working view
One backend truth · three projections · switching mode never restarts execution.
The same Mission - "EU · UK · UAE launch readiness" - seen through the three projections. One left navigation, one composer, one right inspector, one Human Gate card: only the depth of what is shown changes. The Human Gate card ("Approve staged production rollout") is present in every mode: approval is a property of the Mission, not of the view.
Navigation, Missions, the Needs-you queue, the composer with Intelligence Profile and approval policy, and the Progress / Output / Context inspector are identical in all three modes.
FOCUS shows the conversation and the decision. NORMAL adds the execution plan, agents, artifacts and evidence. JARVIS adds the command core, live orchestration, observed build and the task graph. Nothing runs differently underneath.
Every number on these screens (sources, claims, progress, ETA) is synthetic prototype data used to design the surface; efficiency and cost tiles are marked DEMO / SIMULATED. None is a measured result, a benchmark or a commercial commitment.
The face of the platform
LILI is the conversational presence the user talks to. LILI is the face, NEO AI is the platform, Mission Control is the OS view. A luminous presence - never a face, an avatar or a robot.
A conversation that does the work
Chat-first, with visible history and message search - not a console with a chat box. NEO responds with visible progressive work: rationale, evidence, decisions and provenance - never private chain-of-thought.
Progress · Output · Context
A universal inspector exposes the Mission in every mode. Extra depth arrives through the inspector, expandable panes and drill-down - JARVIS never becomes a different application.
Nothing is simulated
No fake agents, progress, tests, evidence, cost or verified state. A capability not yet connected is shown as such - the interface never implies a result that was not produced.
Observe Build
See the work. See the result. Control the release.
When NEO creates an application, a website or a software product, the user does not merely receive code at the end. The user sees the build and the actual running result evolve: code being produced, files changing, build progress, tests, verification, the running preview in the App Viewer at the same revision as the visible code, deployment status - and a Human Gate before every protected release action.
What the user sees
- Actions and rationale summaries - never private chain-of-thought
- Code, files and artifacts as they change
- Build progress, tests and independent verification
- The running preview - interactive, same revision as the code
- Decisions, evidence, provenance and state transitions
- Deployment status and the Human Gate before release
The canonical product principle for everything NEO builds: the user sees the work as it happens, sees the real running result - interactive, at the same revision as the visible code - and controls what is released.
- Blueprint
- Build
- Preview ready
- Connect domain
- Verify
- Human Gate
- Deploy
- Human Gate
- Live
For runnable software the user sees the real running result, interactive, at the same revision as the visible code - never a screenshot or a mockup.
A failed candidate never replaces the previous healthy preview, and the two can never be confused.
One Mission, not a separate app; where permitted, builds and tests run sandboxed on the user's workstation.
NEO Intelligence Profiles
The user chooses an Intelligence Profile - not a model.
NEO is not a model switcher and not an API reseller. The user selects how much intelligence a Mission should mobilize; NEO decides which models, agents, tools, memory and compute deliver it - and how much of that envelope comes from the user's device, NEO, private infrastructure or external models: profile ≠ provider ≠ execution location. Five public profiles form the family.
NEO decides the envelope
fast · efficient
balanced
deep reasoning
maximum intelligence
Mode x Profile - every combination valid
FOCUS + APEX - 15 combinations · zero coupling. FOCUS + APEX and JARVIS + PULSE are both valid.
Mode is visibility. Profile is the resource envelope.
NEO AUTO / NEO PULSE / NEO PRIME / NEO AXIOM / NEO APEX are the public profile names; the one-line descriptions above are indicative of intent, not published specifications. Internal profile identities are stable regardless of naming. APEX means maximum justified intelligence, not indiscriminate compute.
Six dimensions that never collapse - plus one
Personal Context gives NEO read access to what the user chooses to share. It is never model training. Broad computer access never implies training permission.
Training eligibility is a separate, explicit choice - off by default and independent of every context grant. No pre-ticked global consent.
Computer Control (Take Control, Pause, Stop, emergency stop) is distinct from Personal Context. Local Compute Permission is a seventh dimension: GPU use never authorizes reading files, acting or training.
Human authority by design
Autonomy in NEO is bounded by construction
The user sets an approval policy - ASK ME, AUTO or AUTONOMOUS - and that policy can only narrow what is auto-approved. Mandatory Human Gates sit outside every policy: no setting, profile or mode reaches them.
The approval policy only narrows what is auto-approved.
- Production DNS changes
- Publishing to a public domain
- Act-class connector actions
- Release of a deployment
No setting reaches these.
- Connected - a service is linked - is not
- Memory - what NEO retains - is not
- Training consent - off by default - is not
- Authority to act - explicit per action
Gmail connected does not authorize sending.
Bounded autonomy - the rules that no setting can override
Connecting an external service is one click for the user, but NEO runs the shortest valid flow - never skipping OAuth, an operating-system permission, an enterprise admin approval or a health-data consent.
Agents receive capability tools, never credentials. Mandatory gates, act-class connector actions, DNS changes and deployment releases are refused at the platform level; interface enforcement is a courtesy.
Enterprise and regulated customers buy bounded autonomy. Authority, memory, training and connectivity that cannot be conflated are what make an AI organization deployable inside a real company.
Security by design - what the architecture enforces
Agents receive capability tools · never credentials
Mandatory Human Gates refused at platform level · interface enforcement is a courtesy
Every artifact carries lineage · verification status · audit trail
Routing constrained by jurisdiction · security policy · authority before cost
Separations that never collapse: CONNECTED ≠ PERSONAL CONTEXT ≠ MEMORY ≠ TRAINING CONSENT ≠ AUTHORITY TO ACT ≠ LOCAL COMPUTE PERMISSION ≠ COMPUTER CONTROL. A connected DNS provider does not authorize publishing; a connected mailbox does not authorize sending; a GPU made available for inference does not authorize reading files or acting. Execution location never bypasses a Human Gate: a build that completed locally still waits for human authority before release. More intelligence, more agents or more compute never create more authority.
Talk to NEO. Watch the work happen.
- FOCUSThe conversation
- NORMALThe plan and progress
- JARVISThe full AI team in action
Live demo
NEO AI at work: three modes, one Mission
FOCUS, NORMAL and JARVIS are three projections of one Mission state - not three applications. These screens are the NEO AI product design itself, running on synthetic fixture data.
Simple · useful · fast
The conversation and the decision. Everything NEO can do remains available underneath; nothing has to be configured to start.
Expanded tools · execution plan
Adds the execution plan, the active agents, artifact progress and evidence. Nothing runs differently underneath.
Full agent ecosystem · command core
Adds the command core, live orchestration, observed build and the task graph - without ever becoming a different application.
Run a Mission
Watch a Mission run - and release it yourself
One conversation, one Mission, one result. Switch between FOCUS, NORMAL and JARVIS while it runs: the work underneath does not restart. Approve the Human Gate yourself to release it.
YouGive me the executive launch decision and prepare the final package.
LILIWorking on it. Planning the Mission and assigning research, legal, build and verification.
LILIResearch is finished. The live dashboard is being built and independently verified.
Recommendation: CONDITIONAL GO
The package is 72% complete. One approval is required before release.
Approve staged production rollout
LILI recommends approval after verification with a rollback snapshot.
LILIApproved by you. Staged rollout released. Final package delivered with evidence and provenance.
Interactive illustration of the Mission "EU · UK · UAE launch readiness". Every figure is synthetic prototype data, not a live Mission, a measured result or a benchmark. No setting bypasses the Human Gate: the release waits for your approval.
Every device
The same Mission whether you speak, type or switch device
Navigation, Missions, the Needs-you queue, the composer with Intelligence Profile and approval policy, and the Progress / Output / Context inspector are identical in all three modes.
Execution may move between devices and clouds; the Mission remains one. The Mission belongs to NEO, not to the device that started it.
Nothing is simulated in the product: a capability not yet connected is shown as such. The screens here carry their FIXTURE labels for that reason.
An AI team that plans, builds and checks.
- In parallelSpecialists work in parallel.
- VerifiedIndependent verification challenges the result.
- Your approvalYou approve protected release actions.
How a Mission is executed
From one Mission to thousands of coordinated sub-tasks
In practice, NEO can break a complex Mission into tens, hundreds or eventually thousands of coordinated sub-tasks, assign each one to the most suitable capability and permitted execution zone, execute them in parallel, cross-check their results, escalate only the difficult parts to more powerful models and assemble the final verified output.
- Decomposetens · hundreds · thousands
- Assignbest-suited capability and zone
- Execute in parallel
- Cross-checkverify results
- Escalateonly the hard parts · frontier model
- Assemblefinal verified output
Decompose & assign
Each unit of work is matched to the best-suited model, agent, tool, memory or data source and to a permitted execution zone - not to one default provider.
Execute & cross-check
Sub-tasks run in parallel. Their results are cross-checked by independent agents before anything reaches the final artifact.
Escalate & assemble
Only the difficult parts are escalated to more powerful models. NEO then assembles the final verified output.
Where the infrastructure technology becomes strategically important
Objective: intelligence multiplied at a fraction of the compute. Brute force: larger models · more GPUs. NEO x100: orchestration · specialization · parallelism · verification · reuse.
Horizontal axis: compute consumed (GPU · tokens · energy · dollars). Vertical axis: effective intelligence. Cyan: NEO x100. Grey: brute force.
Advanced optimization layers
Spend compute where additional intelligence actually creates value
The architecture is designed around several advanced optimization layers.
Mixture-of-Agents (MoA)
Instead of relying on a single AI response, multiple specialized agents can independently research, reason, build, critique and verify different parts of the same Mission. NEO then orchestrates and synthesizes their outputs. The objective is to obtain stronger system-level intelligence from coordinated specialization rather than simply increasing model size.
- Research
- Reason
- Build
- Critique
- Verify
- Synthesis
Adaptive Model Cascades
NEO can start a task on the cheapest model capable of solving it and escalate only when confidence, complexity, verification or policy requires a stronger model. A simple extraction task should not consume the same compute as difficult strategic reasoning.
Dynamic Capability Routing
NEO does not merely select a model. Its routing layer can select among models, tools, runtimes, data sources, verifiers, human capabilities and execution zones - device, customer-private, NEO-controlled, external - according to quality, latency, cost, jurisdiction, security, availability, device capability, data locality, privacy requirement, network state and energy. Capability, not brand: the router answers what should perform the task and where it should execute, and continuously re-optimizes rather than binding a Mission to one provider or one location.
- quality
- latency
- cost
- jurisdiction
- security
- availability
- Router
- Models
- Tools
- Runtimes
- Data sources
- Verifiers
- Humans
Semantic Compute Reuse
Previously computed, verified and still-valid intelligence can potentially be reused instead of regenerated from scratch. Through semantic caching, persistent memory, artifact lineage and verified intermediate results, NEO is designed to reduce duplicated model inference across users, agents and repeated workflows.
- Requests · users · agents
- Semantic cache · memory · lineage · verified
- Reused · still valid
- Regenerate · only if new
Speculative Parallel Intelligence
For difficult problems, NEO can explore several candidate approaches simultaneously using different agents or models, terminate weak branches early and allocate additional compute only to promising or disputed paths. The objective is to spend compute where additional intelligence actually creates value.
- ✗ terminated early
- ✓ more compute here
- ✗ terminated early
- explore · prune · concentrate
Verification-Driven Compute Allocation
Instead of treating inference quality as a fixed property of one expensive model, NEO can use cheaper models or specialist agents for initial work and independent verifiers to determine where additional reasoning is actually necessary. Compute becomes conditional rather than uniformly expensive.
- Cheap model / specialist agent
- Independent verifier
- Accept · no extra cost
- More reasoning · where necessary
Hybrid API / Private / Self-Hosted Execution
NEO is designed to route workloads across edge and device compute, frontier-model APIs, private enterprise models, specialist systems, NEO-controlled and self-hosted open models depending on economics, latency, privacy, data locality and capability requirements - the NEO Execution Fabric.
- Frontier-model APIs
- Private enterprise models
- Specialist systems
- Self-hosted open models
Memory and Context Optimization
Rather than repeatedly sending an entire historical context to every model invocation, the system is designed to construct the minimum authoritative context required for each task, reducing unnecessary token processing while maintaining provenance and continuity.
Together, these technologies create intelligence-per-compute optimization
The goal is not simply: more AI. It is: more useful intelligence per dollar, per GPU, per token and ultimately per watt of energy consumed.
The optimization envelope
The cheapest model does not automatically win
x100 optimization is constrained by quality. NEO minimizes cost, latency, data movement and unnecessary compute subject to the Mission's required quality, verification, security, jurisdiction, authority, privacy and availability envelope - with execution location, device capability and energy state as additional routing inputs. The best execution path is the least expensive permitted one that satisfies that envelope - never a cheaper or more local path that degrades the result.
- cost
- latency
- unnecessary compute
- GPU time
- energy
least expensive path that satisfies the envelope
- quality threshold
- verification threshold
- security policy
- jurisdiction
- authority · Human Gates
- availability
System-level, not single-model
Decomposition, Mixture-of-Agents, specialist tools, escalation and synthesis are designed to produce stronger effective intelligence than any one model used alone.
Compute where it creates value
Cascades, routing, semantic reuse, speculative pruning and conditional compute keep frontier inference for the parts of a Mission that need it.
Verified before it is delivered
Independent verification, cross-checking, provenance and Human Gates make the result something a business can act on - and define the threshold no optimization may cross.
Three Missions, three execution paths
Low-cost capable model → verification passes → done. Frontier compute never touched.
quality · threshold per MissionCascade escalates on confidence and policy → frontier model for the hard part only → independent verification.
verification · independent · requiredRouting is constrained by jurisdiction, security policy and authority before cost is considered → the compliant path wins even if it is not the cheapest.
policy first · then cost, minimized inside the envelopeCompute before tokens
Not every task needs a frontier model. Not every task needs the cloud.
Before a workload becomes tokens, NEO determines whether a deterministic tool, a local runtime, a database query, a parser, a compiler, a vision system or a specialist model can perform it more accurately and economically. A large part of expensive frontier consumption can be avoided before reasoning even begins: models are used for intelligence, tools for computation, and the user's own hardware for both where sufficient.
- Raw Mission workloade.g. a 500-page document set
- Reuse valid workalready verified · fresh · authorized
- Tools · local / edgeOCR · parsing · dedup · indexing · embeddings · retrieval · compression
- Minimum authoritative contextonly the relevant, permitted subset moves
- Premium reasoningfrontier intelligence where it creates value
- Verify · stop when donedeterministic checks first · escalate on failure
Only the hard / disputed subset reaches frontier compute. Segment widths are not measured.
NEO does not pay frontier-model economics to discover which information deserves frontier-model intelligence. Not every problem should become tokens.
1,000 raw documents → edge / private OCR, parsing, deduplication, indexing, relevance scoring and context compression → relevant authoritative context → premium reasoning → verification → decision-ready output. Sensitive source material does not need to follow every reasoning step into the same execution environment.
"Summarize and organize these files." Parsing, indexing, retrieval and suitable local AI run on the machine; only the portions that need stronger intelligence escalate. The user sees one answer and manages no infrastructure. Open NEO. Ask LILI. Receive the result.
Repository indexing, dependency analysis, compilation, tests, linting and local preview on the user's workstation; architecture reasoning and security analysis through NEO and specialist tools; frontier intelligence only where higher capability is valuable. One continuous Mission - no "local mode" and no "cloud mode".
Valid · fresh · authorized · same owner
Deterministic checks and specialist tools first
Move the 20 KB that matters, not the gigabyte
No compute after the envelope is satisfied
Mission economics is the unit of optimization: a cheap first call that causes five retries costs more than one stronger call, and a premium model used strategically can lower total Mission cost. NEO optimizes cost per verified Mission outcome - not the sticker price of each inference. Compute can also increase when value justifies it: high-stakes Missions may deliberately spend more on independent analyses, deeper verification and red-team reasoning.
Economics of x100
Technology → quality → cost → margin → scale
Better, verified intelligence is designed to raise willingness to pay and enterprise usability. Less frontier compute per unit of value can lower the cost to serve. Proprietary optimization could become a second business sold to the AI industry itself.
Maximize: verified intelligence delivered ÷ total compute economic cost - system-level, measured over time
Total economic compute in the denominator: external API inference, NEO-controlled compute, customer-private compute, edge / user-owned compute (not free - electricity, hardware, attention), specialist tools, orchestration, verification, retries, memory and retrieval, data movement, human review. x100 measures net system efficiency including the cost of NEO itself - a direction for this ratio, not a measured multiple. Intelligence per NEO cloud dollar is a secondary internal KPI, never the canonical metric.
- Technologyx100 layers
- Qualityverified output
- Costless frontier compute
- Margingross-margin potential
- Scaleoperating leverage
Better verified intelligence → higher willingness to pay · less frontier compute per unit of value → margin as volume grows.
Where savings can arise - conceptual allocation of 100 work units
Most work never reaches the most expensive model - escalation is the exception, not the default.
Every task does not pay frontier prices
Low-cost model for simple extraction; mid-tier for structured analysis; frontier only for strategic reasoning - and only when the confidence or policy check fails. Escalation is the exception, not the default.
Do not recompute verified work
Mission 1 verifies "Company X is incorporated in jurisdiction Y" and stores it with lineage and freshness. Mission 2 asks the same fact: retrieved and revalidated when necessary, not recomputed. Less duplicated inference means lower unit-cost potential.
The infrastructure thesis
AI scaling creates a major infrastructure challenge
Frontier systems require large quantities of GPUs, electricity, datacenter capacity and capital, and inference economics become a central constraint as AI adoption scales.
NEO's thesis is that a significant part of the next efficiency breakthrough will come from using models more intelligently at the system level and from orchestrating compute already distributed across user devices, enterprise infrastructure, NEO infrastructure and external AI systems - an intelligence and compute orchestration layer, not merely a model router.
Vertical axis: inference cost. Horizontal axis: AI adoption. Grey: without optimization. Cyan: with NEO x100. Conceptual, not measured.
If NEO can demonstrate materially better:
- intelligence per-dollar
- intelligence per-GPU
- intelligence per-token
- intelligence per-watt
while maintaining or improving output quality, then the underlying NEO technology becomes valuable independently of the NEO application itself.
A major second strategic opportunity: NEO as infrastructure for the AI industry
NEO's orchestration and compute-optimization technology could ultimately be licensed or deployed as infrastructure for major AI companies, cloud providers, enterprises and model developers seeking to reduce inference cost, GPU utilization and energy consumption.
Applications & customersNEO SaaS · vertical products · enterprise deployments
NEO intelligence & compute orchestration layerrouting · MoA · cascades · reuse · verification · memory · provenance · edge allocation · privacy policy
Global intelligence & compute pooluser devices · customer private · NEO compute · self-hosted · GPT · Claude · Gemini · Grok · Kimi · Qwen · DeepSeek · specialists
Models remain replaceable · the NEO layer compounds in value.
Potential customer categories, appropriately qualified: enterprises, cloud providers, model developers, AI platforms and infrastructure operators. Organizations operating at the scale of OpenAI, Anthropic, Google, Microsoft, Amazon or NVIDIA illustrate the type and scale of participants in this ecosystem - they are named as examples of the ecosystem, not as commercial prospects. NEO does not replace foundation models; it could make models and compute significantly more efficiently orchestrated. Additional proprietary optimization techniques within the NEO x100 R&D roadmap are intentionally not part of this external description.
More intelligence from every unit of compute.
- ReuseReuse verified work.
- RouteRoute by task. Quality and policy first.
- EscalateEscalate when needed.
NEO Edge Compute
Cloud when needed. Local when sufficient. NEO decides.
NEO is designed to use not only external AI APIs and NEO-controlled infrastructure, but also - where technically available, permitted and beneficial - the compute and AI capabilities already present on the user's own device. Conventional cloud AI runs User → Internet → Cloud model → Result. NEO adds User → NEO → local device intelligence → Result when the device is sufficient, and escalates only the parts of a Mission that need stronger intelligence. The user still chooses an Intelligence Profile, never a model, a vendor or a location.
what · where · how much · verified how
user-owned · authorized
- Apple Silicon · Foundation Models
- NVIDIA · AMD · Intel · Qualcomm
- CPU · GPU · NPU
- local models · local tools · OCR
NEO infrastructure
- NEO cloud · NEO models
- private / self-hosted deployments
- customer-private compute (hybrid)
- verification infrastructure
frontier · specialist
- GPT · Claude · Gemini · Grok
- Kimi · Qwen · DeepSeek
- specialist systems
- future frontier models
- Verification
- Synthesis
- User
Local when sufficient · quality-gated · policy-first · cloud-escalated. Named vendors are ecosystems NEO is designed to integrate, not dependencies or partners.
Potentially fewer external inference calls
Selected workloads remain local where policy requires or allows
Suitable workloads execute close to the user
Selected capabilities continue without continuous cloud dependence
User-owned compute can reduce NEO's direct cloud cost-to-serve
Canonical principles
Authorized local resources are used where they meet the Mission envelope
NEO never saves compute by degrading the required result - a local model does not win because it is already paid for
Security, privacy, jurisdiction and authority precede cost; the cheapest route never overrides the permitted execution zone
Stronger, private or cloud intelligence whenever the local path does not satisfy the envelope
Apple, Windows and Linux are execution ecosystems, not separate products; no Mac is required
Local resources are permissioned, bounded, observable and revocable - never harvested
Execution may move between devices and clouds; the Mission remains one
Models, runtimes and hardware backends remain replaceable behind one capability abstraction
Limitations and dependencies: heterogeneous hardware, local model quality, RAM / VRAM, battery and thermal constraints, driver and OS restrictions, model licensing, endpoint security and platform-vendor policies. Mitigation is architectural: NEO detects capabilities, benchmarks execution paths, applies policy and quality thresholds and escalates whenever local execution does not satisfy the Mission envelope. A user's device serves that user's own Missions; no cross-customer compute network is implied. User-owned compute is not free compute: it consumes electricity, hardware and attention - what it can reduce is NEO's direct external inference and centralized infrastructure spend.
Cross-platform by design
NEO adapts to the compute available. The user does not adapt to NEO.
Apple is one important implementation path: on supported Apple devices NEO Edge Compute is designed to use Apple Silicon and, where available, the Foundation Models framework and other on-device Apple capabilities for suitable local workloads. Windows and Linux are first-class targets on the same architecture: NEO detects CPU, GPU, NPU, memory, storage and supported runtimes and builds a Device Compute Profile that tells the router what can, should and cannot run on that machine. A phone, a laptop and a GPU workstation receive different local envelopes; the Mission quality target is the same.
macOS · iPhone · iPad
- Apple Silicon CPU / GPU / Neural Engine
- Foundation Models framework (where supported)
- Core ML · MLX · Vision framework
- permitted local / open-weight models
workstations · AI PCs · laptops
- NVIDIA · AMD GPUs
- Intel CPU / GPU / NPU
- Qualcomm and other NPUs
- supported local inference runtimes
workstations · GPU servers
- GPU workstations · GPU servers
- CPU · supported accelerators
- supported inference runtimes
- developer and enterprise environments
NEO Edge Runtimedevice capability engine · local resource governor · local model manager · sandboxes · provenance
NEO Routercapability, not brand
Any device · any model · any compute · one NEO
Architectural direction; supported hardware / runtime combinations only; no Mac required.
- Device: laptop · workstation · phone
- CPU · GPU · NPU / Neural Engine
- RAM · VRAM · storage · OS
- supported runtimes · local models
- battery · thermal · network state
- security / policy constraints
- safe benchmark, not spec sheets
what can · should · cannot run here
CPU · GPU · battery · foreground first. Resource use is workload-aware, bounded, interruptible and never a substitute for the user's own use of the device.
Locally: OCR, embeddings, classification, routing, intent detection, privacy pre-classification, small models, preprocessing. Larger Missions use the device as the interface while execution occurs elsewhere.
Additionally: local language and vision models for summarization, extraction, translation, context compression, local retrieval over Personal Context, local agents where appropriate.
Potentially: larger local models, parallel agent workloads, document corpora, coding workloads, builds and tests, vision and media processing, specialized NEO models - subject to benchmarking.
Edge Compute enhances NEO; it is not a prerequisite. A device without local AI support runs NEO fully through NEO-controlled and approved external resources. The NEO Edge Runtime is the trusted local execution layer - signed, sandboxed, least-privilege, with integrity-verified models - and it is not a remote-code-execution channel: every local capability is defined, permissioned, auditable and revocable. Named technologies are current third-party ecosystems NEO is designed to integrate where supported; exact framework support is verified per release.
NEO Execution Fabric
Which capability should do the work - and where should it execute?
Dynamic Capability Routing answers two questions at once. Every unit of work is the intersection of three envelopes - Intelligence (how much reasoning and verification, set by the profile), Policy (where data and actions may go) and Capability (what resources exist now, from the Device Compute Profile to reserved cloud capacity). NEO optimizes cost, latency and energy only inside that valid space, across four normalized execution zones.
- NEO Mission
- Mission decompositionNEO Router
how much reasoning · verification
where data and actions may go
what resources exist now
valid execution space ∩ lowest total cost
Where may this work run? Pick a workload.
Any zone.Edge, Customer Private or NEO Controlled.Edge only.NEO Controlled or External under data minimization.
Apple · Windows · Linux
CPU · GPU · NPU · local models
workstations · servers · private cloud
private models · enterprise data
NEO cloud · NEO models
managed runtime · verifiers
frontier APIs · specialist systems
approved providers only
No compliant path: NEO fails safely, never routes around policy.
- Execution
- Verification
- Synthesis
- Verified outcome
One Mission state · one policy plane · one verification framework · one provenance model - multiple intelligence and compute resources.
- Task → local capable?
- Local execution
- Verify (deterministic first)
- Threshold met? Yes → done
- No → escalate to stronger resource → verify
- Task → data / privacy classification
- public · internal · confidential · regulated
- Permitted execution zones
- Capability routing inside those zones
- Then cost · latency · energy optimization
Order of precedence
- 1legal · policy · authority
- 2security · privacy
- 3quality · verification
- 4capability · availability
- 5latency · cost · energy
Public research: any zone. Confidential corporate document: Edge, Customer Private or NEO Controlled. Device-only document: Edge only. Frontier reasoning: NEO Controlled or External under data minimization. No compliant path → NEO fails safely, never routes around policy.
Move only the data the selected path needs. Sensitive sources are processed and minimized in their zone so only permitted derived context reaches stronger models - subject to policy; local preprocessing does not by itself make external processing compliant.
Deployment patterns describe how NEO is deployed; execution zones describe where a task runs. Edge participates in every pattern - managed, private cloud, hybrid, self-hosted, sovereign - with identical Mission, policy, audit and NEO Credits semantics.
Model sovereignty roadmap
API-first → hybrid → NEO-optimized model and execution layer
NEO does not own a frontier foundation model today and does not claim to. The staged strategy reduces structural dependence on any single provider while preserving access to whichever external frontier systems are strongest. The objective is not ideological independence - it is optimal intelligence, resilience and economics.
Multi-provider API orchestration
Leading external AI systems used through APIs where they provide the best capability and economics: GPT, Claude, Gemini, Grok, Kimi, Qwen, DeepSeek, open-weight and specialist models - while edge and local capabilities begin handling selected workloads where supported. NEO remains model-agnostic; no provider is structurally mandatory.
Hybrid execution
Frontier APIs continue while selected open-weight and private models are progressively deployed on infrastructure controlled by NEO: cloud, private enterprise and self-hosted execution, specialist models, routing between external and internal compute, workload-specific economics. Open-weight deployment is part of this strategy; no third-party model IP is owned.
NEO-optimized model and execution layer
A rising share of suitable workloads handled through NEO-optimized, private and edge execution: fine-tuned, distilled, domain-specialized and smaller task-specific models, private and self-hosted open-weight models, reusable intelligence, routing intelligence and evaluation-driven model selection. External frontier APIs remain available.
Share of workloads: conceptual, not a forecast. External systems are inputs that remain replaceable · no provider is structurally mandatory · no ownership of third-party model IP is implied.
Long-term direction: maximum intelligence + minimum structural provider dependency + optimal economics + resilience - model sovereignty and compute sovereignty together, without isolation. Models remain replaceable. Compute providers and hardware backends remain replaceable. The system compounds.
Deployment architecture
One system, five deployment patterns - edge participates in all of them
The same NEO system is designed to run as a managed service, in a dedicated private cloud, in hybrid form with a NEO control plane over customer resources, self-hosted on private infrastructure and, where required and legally supported, in sovereign high-control environments - with the same Mission, policy and audit semantics everywhere.
| Pattern | Where data and compute live | Who operates | Typical driver | Status |
|---|---|---|---|---|
| Public / managed | NEO-managed environment | NEO | speed · simplicity · individuals and SMEs | Target |
| Private cloud | dedicated customer environment | NEO for the customer | data isolation · enterprise procurement | Roadmap |
| Hybrid | NEO control plane + customer resources | shared | latency · residency · existing infrastructure | Roadmap |
| Self-hosted | customer private infrastructure | customer | regulatory perimeter · model sovereignty | Roadmap |
| Sovereign / high-control | national or high-control environment | customer or mandated operator | government adoption · legal framework | Roadmap |
Why it matters: data residency · privacy · latency · regulatory requirements · security · model sovereignty · enterprise procurement · government adoption. The O7 constraint: the platform runs in a client VPC and on-premise with the same Mission, policy and audit semantics. Residency and sovereignty constraints are written before a cloud provider is chosen - no provider decision is implied here.
Residency, privacy, latency, regulation and security are procurement conditions for enterprises, regulated institutions and public-sector buyers.
Mission state, Intelligence Profiles, Human Gates, policy, verification, provenance, memory boundaries and NEO Credits metering behave identically across patterns and execution zones; only where the compute and data live changes.
NEO-managed SaaS is the target delivery model. Private cloud, hybrid, self-hosted and sovereign patterns are architecture-supported roadmap options; none is stated as production-ready here. Edge Compute is an execution layer available to every pattern, not a sixth topology.
An AI team for everyone.
- EntrepreneursBuild the business
- StudentsLearn and create
- ProfessionalsDeliver the work
An AI organization
An AI organization, not an AI assistant
At the application level, NEO can be understood as an AI organization rather than an AI assistant. The organization is logical; its execution is heterogeneous - a document agent may work on the user's device, a financial model in deterministic local computation, a research agent through external tools, a strategy agent through frontier reasoning and a verifier through an independent route, all under one Mission.
Dozens of specialized agents can research markets, competitors, regulations, scientific literature, financial information and technical sources, challenge one another's conclusions, verify evidence and produce a board-ready strategic report.
Architecture, coding, testing, security, debugging, documentation and review agents can work simultaneously on the same software Mission, while NEO dynamically allocates different models and tools to each part of the build.
Agents can analyze companies, individuals, corporate structures, wallets, blockchain transactions, domains, documents, sources and relationships, reconstruct timelines, identify contradictions and generate evidence-backed investigation reports.
Specialized agents can analyze legislation, contracts, cases, regulatory requirements and evidence, while independent verification layers challenge claims before they reach the final artifact.
NEO can combine research, financial modelling, scenario analysis, risk assessment, source verification and independent critique.
NEO can coordinate intelligence across corporate documents, APIs, databases, workflows and internal systems rather than operating as an isolated chat interface.
1 model → 1 answer
Many capabilities → verified output
In a traditional assistant experience the user primarily interacts with a provider's assistant and model stack. In NEO, the Mission is explicitly orchestrated across heterogeneous models, agents, tools, memory, data, verification and execution environments.
The digital workforce
An entire digital workforce organized around an objective
Where a traditional assistant may provide one answer, NEO can organize an entire digital workforce around an objective - all coordinated through the same Intelligence Operating System.
Mission
- Research Agents
- Legal Agents
- Build Agents
- Evidence Agents
- Financial Agents
- Security Agents
- Risk Critics
- Verification Agents
- Specialist Agents
Why can this be more powerful than using ChatGPT or Claude alone?
Because there is no reason to assume one model - or one place to run it - will always be the best at every component of every task.
- One model may be strongest at deep reasoning.
- Another may be better for coding.
- Another may offer better latency.
- Another may be dramatically cheaper for extraction or classification.
- A local model on the user's own device may be the right choice where privacy or proximity matters.
- A specialist tool may outperform every general-purpose LLM for a particular operation.
| Deep reasoning | Coding | Latency | Cost (extraction) | Privacy / local | Specialist op. | |
|---|---|---|---|---|---|---|
| Model A | ★ | |||||
| Model B | ★ | |||||
| Model C | ★ | |||||
| Local / edge model | ★ | ★ | ||||
| Specialist tool | ★ | ★ |
NEO combines the best capability and the best execution location for each component instead of forcing one model to do everything.
Multiple products - one core
A new vertical is not a new company
A new vertical reuses the common NEO platform - orchestration, identity, billing, models, agents, memory, governance, verification, compute, the Execution Fabric with Edge Compute, and enterprise integration - rather than requiring a new platform from scratch. Every vertical therefore benefits from the same local / private / cloud orchestration; no vertical builds its own edge stack. Incremental work concentrates in domain configuration, data and connectors, compliance, product adaptation and go-to-market.
CORE
- orchestration
- identity
- billing
- models
- agents
- memory
- governance
- verification
- compute
- integration
- Software Factory
- Company Builder
- Fintech Studio
- Payments
- Sport & Wellness
- RWA Studio
- Hospitality & Nightlife
- Micro-Credit
- Trading & Treasury
- Health & Care
- Legal Ops
- Travel
- NEO Intelligence
Inner ring = shared services reused by every vertical · outer ring = revenue surfaces · 13 verticals.
Shared services amortize
Shared infrastructure amortizes across verticals; each additional product carries a lower fixed-cost burden - the source of operating leverage. Edge examples only: Software Factory - local builds and tests; Legal - private document processing; Finance - local deterministic computation; NEO Intelligence - local evidence preprocessing.
Same product, configured per industry
NEO AI is one product. Every vertical is the same Intelligence Operating System configured for an industry - never a separate product or company.
Not dependent on one vertical winning
Three pilot-anchored verticals exist today; the others are built on the same core when demand is proven.
Market architecture
Individuals, organizations and, where applicable, sovereign users
NEO AI is not an enterprise-only platform. Like a general-purpose AI assistant, the core Intelligence Operating System is usable directly by private individuals and professionals - and the same core serves SMEs, enterprises, regulated institutions and, subject to the relevant product and legal framework, government and sovereign deployments. Every vertical is a configuration of the same system, exposed to each customer class with different permissions, data access, connectors, compliance, compute envelope - including edge permissions, approved Compute Nodes, private infrastructure and external-provider eligibility - pricing, deployment, authority, security and tools.
CORE
- B2Cconsumer experience
- B2G · Sovereignwhere applicable
- B2Benterprise deployment
Individual and professional subscriptions · NEO Credits
- individuals
- professionals
- creators
- developers
- researchers
- investors
Vertical products · enterprise and private deployments
- SMEs
- enterprises
- professional firms
- financial institutions
- regulated organizations
- large groups
Sovereign programs · subject to product and legal framework
- government
- public sector
- sovereign programs
The same vertical ecosystem · one shared core · no duplicated technology stack
What differs by customer class: permissions · data access · connectors · compliance · compute envelope · pricing · deployment · authority · security · tools.
A private user can use NEO Legal capabilities for research and analysis; an enterprise legal environment adds private data, enterprise connectors, governance and dedicated infrastructure.
A private investor can use Finance and Research capabilities; a financial institution receives enterprise-grade data, controls and private deployment. The same holds for compute: a consumer's capable computer, an employee's managed laptop and an enterprise GPU pool are all authorized resources for their own Missions - never for other customers.
Bounded by lawful access, provider licensing, data rights and Human Gates. A consumer never gains access to restricted enterprise, regulated or licensed datasets because the same vertical exists.
NEO Intelligence
Not a dashboard. Not a database. An intelligence layer.
NEO Intelligence is designed as the orchestration and reasoning layer above intelligence providers: OSINT, corporate registries and UBO, blockchain and wallet intelligence, AML and sanctions screening through licensed providers, adverse media, documents, relationships, timelines and evidence - resolved into one provenance-tracked, evidence-backed report under a human approval gate. NEO orchestrates screening results; it is not an adjudication authority. Edge and customer-private compute can preprocess large evidence sets - OCR, entity extraction, deduplication, timeline and relationship candidates - before cross-source reasoning; licensed external sources remain governed by provider terms and lawful access. Local compute changes where processing happens, not the user's legal rights to data.
- Sourceregistry · wallet · document · media
- Agentextract · scan
- Evidenceentity resolved
- FindingA · 0.9 - B · 0.7 - C · 0.55
- Human Gate
- Evidence-backed report
Confidence is never presented as certainty.
- OSINT
- Registries & UBO
- Blockchain & wallets
- AML & sanctions screening
- Adverse media
- Cyber threat
- Entity resolution
- Timelines
- Evidence-backed reports
- Banks and regulated institutions - KYC/KYB, enhanced due diligence, transaction context
- Law firms - litigation support, asset tracing, evidence assembly
- Corporates and funds - counterparty, M&A and investment due diligence
Lawful access only under applicable law and contract · connector roadmap subject to provider access and licensing · NEO does not own or resell third-party databases · human review on high-impact conclusions · confidence scores never presented as certainty · no law-enforcement powers, no final legal or sanctions determinations.
Same orchestration · verification · memory
Orchestrates sources · never replaces them
The value sits above the models.
- One platformTwo revenue engines.
- Individuals & teamsA subscription that includes NEO Credits. A plan that grows with use.
- EnterprisesPaid pilot, annual contract, expansion.
Business model
Several layers, two businesses
Layers 1 to 4 are the Intelligence Operating System used by customers. Layer 5 is the potential licensing of NEO's compute-optimization infrastructure to the AI industry itself.
NEO SaaSSubscriptions for individuals, professionals and companies - B2C and B2B on the same core; premium Intelligence Profiles and capabilities where approved.
Vertical Intelligence ProductsDedicated environments for areas such as Research, Legal, Finance, Software Development, NEO Intelligence/OSINT and enterprise operations.
Usage & Compute · NEO CreditsThe customer buys NEO Credits, never the edge / private / NEO / external mix behind them. Subscriptions include a level of usage, with additional intelligence/compute capacity available when required. The customer-facing usage unit is NEO Credits; provider-native tokens stay internal metering only.
Enterprise DeploymentsPrivate infrastructure, proprietary company data, private models, advanced security, dedicated environments and custom integrations.
NEO Infrastructure TechnologyPotential licensing of NEO's intelligence-and-compute orchestration (incl. hardware-aware routing and edge / cloud placement), routing and compute-efficiency stack to enterprises, cloud providers, model companies and AI infrastructure operators.
An Intelligence Operating System used by customers
SaaS · Verticals · Usage · Enterprise
An intelligence/compute optimization infrastructure layer used by the AI industry itself
Compute Optimization Infrastructure Layer · optionality
NEO Credits
The commercial usage unit
Customers buy and consume NEO Credits. Provider tokens, GPU seconds and edge execution remain internal metering only. The chain from execution usage - edge, private, NEO, external, tools, reuse - to internal resource accounting, to a normalized NEO resource, to NEO Credits stays separated, so pricing follows the intelligence delivered, not any provider's token bill or the location where work ran.
- Provider usagetokens · calls · runtime - internal
- Provider costwhat each provider bills - internal
- Normalized NEO resourceone unit across providers - internal
- NEO Creditsthe customer-facing unit
Subscriptions include NEO Credits · Intelligence Profiles consume them at different envelopes · provider-native tokens never reach the customer.
How it reads in the business model
Subscriptions include a level of NEO Credits; additional Credits are billed on top when a customer needs more intelligence or compute capacity.
Envelope, not provider
NEO AUTO, PULSE, PRIME, AXIOM and APEX consume Credits at different envelopes. The customer decides how much intelligence a Mission mobilizes - never which provider is paid.
x100 accrues to NEO
Because the customer unit is decoupled from provider tokens, every gain in intelligence-per-compute - routing, cascades, reuse, verification, edge and user-owned compute, specialist tools, context compression - can lower NEO's cost to serve without changing what the customer buys. Whether local execution changes Credit consumption is a future commercial decision; Credits are never a direct local-compute meter.
NEO Credits are an internal commercial usage unit, not a cryptocurrency, transferable token or financial instrument - and not tied to any provider's tariff.
Provider costs can move; the customer unit and its meaning do not.
Credits meter intelligence mobilized, not a guaranteed quantity of output.
Credit allowances and per-Credit prices are commercial parameters set at launch of each plan; they are not stated here.
Compounding intelligence
Every Mission can improve system-level orchestration - without automatically using customer content for model training
Mission execution data, evaluation signals, routing performance, outcome telemetry, verification performance and reusable verified intelligence are designed to improve how NEO allocates resources. Training-eligible data is a separate category, governed by explicit policy, consent and lawful basis. The two are never the same thing.
INTELLIGENCE
- 1 · Mission
- 2 · Execution
- 3 · Outcome
- 4 · Evaluation
- 5 · Routing knowledge
- 6 · Better allocation
- 7 · Quality · cost · latency
- 8 · More Missions
Mission → execution → evaluation → model, tool, hardware and runtime knowledge → better allocation → more Missions. No weights change.
Training eligibility → explicit policy, consent and lawful basis → only eligible data enters applicable improvement or training processes. Help Improve NEO is off by default and independent of every context grant.
Privacy, enterprise adoption, regulatory defensibility and product trust depend on this separation - and the routing and evaluation knowledge that does accumulate belongs to the NEO system layer - which model, hardware and runtime wins which task - without customer content.
| Signal | What it improves | Contains customer content? | Enters model training? |
|---|---|---|---|
| Mission execution data | orchestration and task-graph design | bounded to the Mission | no - unless training-eligible by explicit consent |
| Evaluation signals | quality scoring of models, agents and tools | no | no |
| Routing performance | which resource wins which task, incl. hardware class and zone | no | no |
| Outcome telemetry | cost · latency · success rates | no | no |
| Verification performance | where extra reasoning is actually needed | no | no |
| Reusable verified intelligence | semantic reuse with provenance and freshness | scoped to its owner and permissions | no |
| Training-eligible data | applicable improvement or training processes | only what policy and consent allow | yes - explicit policy · consent · lawful basis |
Customer confidential information is never automatically used to train models.
Where value compounds
Models remain replaceable. The NEO architecture compounds.
NEO can combine these resources rather than forcing one model to perform everything. The result is the potential for greater effective intelligence at the system level than any individual model used in isolation, while simultaneously reducing unnecessary use of the most expensive compute. The long-term strategic value is therefore that models remain replaceable, while the NEO intelligence architecture compounds in value.
MOAT
- Orchestration
- Routing
- Agent coordination
- Accumulated system intelligence
- Evaluation data
- Memory
- Compute optimization
- Verification
- Enterprise integrations
- Workflows
- Provenance
Replaceable
GPT, Claude, Gemini, Grok, Kimi, Qwen, DeepSeek, private and open models can be swapped as the market evolves - and so can API providers, clouds, runtimes, GPUs and NPUs; if one deteriorates, NEO routes around it.
Compounding
Orchestration, routing, model and hardware evaluation (the NEO Capability Graph), edge / cloud allocation, runtime abstraction, privacy-aware routing, reuse, memory, verification, provenance, integrations, workflow knowledge and vertical configurations accumulate with every Mission. The resources improve; NEO learns how to use them better.
- Todayorchestration · verification · gates
- System learningevaluation · routing · reuse
- Model sovereigntyprivate · fine-tuned · hybrid
- Increasingly general system-level intelligenceresearch direction · no date
Long-term research direction: a direction of travel, not a roadmap commitment, not a demonstrated result. Increasingly general system-level intelligence could emerge from orchestration, agents, memory, verification, self-improvement and model sovereignty - a research ambition, not a current capability, not a demonstrated result, not a guaranteed outcome.
The industry measures intelligence at the model. We study the system.
- 8research pillars
- 16strategic studies
- 8industry reports
Research
Research that becomes the platform
Advancing the frontiers of artificial intelligence through rigorous, institutional-grade research across eight strategic pillars including AI + Crypto and AI + Blockchain. Mickael Mosse, Group CEO of My NEO Group, publishes strategic studies, industry reports and institutional analysis focused on enterprise artificial intelligence, digital infrastructure, regulated industries and the transformation of global business models at mickaelmosse.ai.
NEO AI is the enterprise intelligence operating system designed to operationalize the research published here. Each research pillar informs a specific NEO AI capability - from agentic orchestration to private AI deployment and governance automation. The platform translates institutional research into executable enterprise workflows.
- AI Orchestration
- Private AI
- Governance
- Mission Control
Research pillars
Eight pillars, one intellectual framework
Research pillars define the long-term intellectual framework. AI Insights and Publications provide the sector-specific studies, industry reports and strategic analysis connected to each pillar.
RESEARCH
PROGRAM
- Agentic AI Systems
- AI Operating Systems
- Enterprise Intelligence
- Financial AI
- Healthcare AI
- Artificial General Intelligence
- AI + Crypto
- AI + Blockchain
Agentic AI marks the transition from passive AI tools to systems capable of planning and executing multi-step tasks. The opportunity is substantial: agents can reduce operational friction, coordinate workflows and extend organizational capacity. The risk is equally material: once AI systems can take action, enterprises must govern permissions, accountability, security, auditability and human intervention with far greater discipline.
Thesis: agentic AI will be adopted seriously only when enterprises solve the control problem.
The next stage of enterprise AI will not be defined by isolated models. It will be defined by orchestration. An AI Operating System is the institutional layer that connects models, data, agents, workflows, permissions, audit trails and governance into a controlled enterprise environment. Without this layer, AI adoption remains fragmented and difficult to scale.
Thesis: enterprise AI will fail to scale if it remains a collection of disconnected tools.
Agentic AI Systems
Autonomous AI systems capable of reasoning, planning, tool use and workflow execution under controlled enterprise governance.
AI Operating Systems
Enterprise AI orchestration layers connecting models, data, tools, workflows, permissions, auditability and human decision-making.
Enterprise Intelligence
AI-native architectures that convert organizational data, workflows and institutional knowledge into scalable decision intelligence.
Financial AI
Machine intelligence for banking, risk, fraud detection, compliance, capital allocation, market intelligence and digital financial infrastructure.
Healthcare AI
Clinical workflow intelligence, patient engagement, administrative efficiency, medical AI governance and responsible health-system deployment.
Artificial General Intelligence
Research into increasingly general AI systems, frontier capabilities, autonomy, alignment, institutional preparedness and long-term risk.
AI + Crypto
Machine intelligence applied to digital asset markets, crypto due diligence, lawful intelligence workflows and investor protection frameworks.
AI + Blockchain
AI-enhanced blockchain infrastructure, smart contract intelligence, decentralized governance and institutional-grade distributed systems.
The questions an AI Operating System has to answer
- Which models are used for which tasks?
- Which users and agents can access which data?
- Which tools can an AI system call?
- Which workflows require human approval?
- How are outputs validated?
- How are decisions logged?
- How are risks monitored?
- How is value measured?
- How is compliance enforced?
Founder thesis
From model intelligence to system intelligence
Original writing on the ideas behind NEO AI. These are positions held by Mickael Mosse, not institutional publications of the Research Program. The industry measures intelligence at the model, and for most enterprise work that is the wrong boundary.
A model benchmark is a well-posed question: given this input, is the output acceptable. Real work is not posed that way. Assess whether a counterparty is an unacceptable risk, given twelve documents in three languages, a policy, a regulatory perimeter, a deadline, and an obligation to justify the answer to somebody who was not in the room.
A stronger model improves the raw material for all four requirements and supplies none of them.
RESULT
- Attributablea source, not a generation
- Under an authoritywho was entitled to decide
- Boundedfewer actions than components can perform
- Recordedwhat was known, verified, approved
Model improvement moves one term. It raises what a single call can do. It does not touch attribution, authority, boundedness or the record. So as models get better, the unsolved parts become a larger fraction of what remains, not a smaller one.
When a model does 40 percent of the reasoning you need, the reasoning gap dominates and governance looks like overhead. When it does 90 percent, what stands between you and a usable result is almost entirely those four properties.
A more capable model gets entrusted with longer chains of consequential action, and the cost of an unattributable or unauthorized action rises with the length of the chain. Capability and required governance move together.
If a single frontier model reliably decomposes a task, invokes tools under externally enforced permission, grades evidence by distinct origin, verifies itself through something that is not itself, and emits a record an auditor accepts, then the layer is a temporary artifact of current limitations. I watch that condition deliberately. Progress against it is progress against me.
Precision is a product feature
I run the Program on a rule that is unusual for a company at this stage: a claim about what we do has to be falsifiable, and the status of every capability we describe has to be published, including the ones that do not exist.
Small buyers rarely check. The large, regulated, long-contract buyers who are the actual market check everything. A corpus where a reader can tell established fact from architectural interpretation from design intent is a corpus a serious buyer can use.
When the Program's own red team counted claim tags across our standards, five contained none at all and one had overstated its own count by roughly fourfold. We published that. A standard that exempts its author is worth nothing.
The question is not whether the system may act. It is who was entitled to let it.
Most AI governance material answers one question: is this action permitted. That is necessary and it is half the problem. The other half is entitlement. When an audit asks why an autonomous system did something consequential, "the policy allowed it" is not an answer.
A system that records permission without entitlement has produced a log that looks like accountability and is not. That is why NEO separates the two, structurally. It is the difference between a system an organization can defend and one it can only explain.
I hold this as a design position rather than as a proven result, and the Program's papers tag it accordingly.
Recommended citation: Mosse, M. (2025). "The System Intelligence Thesis." mickaelmosse.ai/thesis/system-intelligence
NEO AI Research Program
The corpus behind the architecture
The technical position is published at the Research Program with claim classes and a capability register. The NEO AI Research Program publishes its corpus at neoai.myneogroup.com/research. Below is a selection from it, with why each one matters. Papers are not reproduced here.
NEO-AI-P-001FORTHCOMING
Introducing NEO AI: Intelligence Orchestrated
The position statement, and the one that discloses how much of the platform exists.
NEO-AI-P-007FORTHCOMING
A Note on Responsible AI Capability Claims
The standard the rest is written under, and the one I would read first if I were assessing us.
NEO-AI-R-009FORTHCOMING
From Model Intelligence to System Intelligence
The centre of the argument. Identifier requested and not yet allocated, so cite it by title.
NEO-AI-TN-001FORTHCOMING
Bounding the Autonomy Budget
Short, and the most directly useful to anyone building.
NEO-AI-TN-002FORTHCOMING
Detecting Circular Corroboration in an Evidence Chain
Five documents restating one press release are one source.
NEO-AI-TN-003FORTHCOMING
What Belongs in a Mission Event
What a governance record has to answer, and what it must never contain.
FULL PAPERFORTHCOMING
Mission Control: A Governance Architecture for Autonomous Work
The paper behind the human-authority thesis.
Industry reports
Eight sectors, eight institutional reports
Institutional-grade research publications, white papers, and industry reports covering the full spectrum of enterprise AI transformation.
| Sector | Report | What it examines |
|---|---|---|
| AI + Banking | Strategic Outlook | How AI is reshaping credit intelligence, compliance, fraud, customer relationships and capital allocation in banking. |
| AI + Blockchain | Institutional Adoption Report | How AI and blockchain converge around tokenization, programmable settlement, digital identity and verifiable execution. |
| AI + Crypto | Stablecoins, Compliance and AI Agents | How stablecoins, MiCA, AI surveillance and machine payments are shaping the next digital asset market structure. |
| AI + Hedge Funds | Alternative Data and Machine Intelligence | How AI is changing hedge fund research, alternative data, portfolio analytics, risk controls and investor due diligence. |
| AI + Healthcare | Clinical Workflow Intelligence | How AI can improve documentation, triage, patient engagement, clinical productivity and healthcare operations under strict governance. |
| AI + Hospitality | Predictive Guest Experience and AI-First Hotels | How AI is transforming hotel discovery, guest personalization, revenue management, loyalty and service operations. |
| AI + Real Estate | Asset Intelligence and AI Infrastructure | How AI is changing real estate asset management while creating new demand for data centers, power, cooling and digital infrastructure. |
| AI + Government | Sovereign AI and Public Administration | How governments can use AI to improve state capacity while managing legitimacy, sovereignty, transparency and public trust. |
Strategic studies
AI Insights: sixteen executive-level studies
Executive-level strategic studies on artificial intelligence, institutional transformation and the sectors being reshaped by enterprise AI.
AI + BANKING
The Intelligent Balance Sheet: How AI Will Redesign Banking From Compliance to Capital Allocation
Artificial intelligence will not simply make banks more efficient. It will reshape how banks understand risk, allocate capital, detect financial crime, serve clients and satisfy regulators.
AI + BANKING
From Digital Banking to Autonomous Banking: The Next Competitive Moat
Digital banking was about access. Autonomous banking is about intelligence: AI-native decision systems that can predict, monitor and respond under strict governance.
AI + BLOCKCHAIN
The Programmable Trust Layer: Why AI and Blockchain Are Converging in Institutional Finance
AI creates intelligence. Blockchain can create verifiable execution and programmable records. Their convergence matters for institutional finance.
AI + BLOCKCHAIN
AI-Native Tokenisation: The Next Operating System for Real-World Assets
The deeper opportunity is to create AI-readable, programmable and continuously monitored asset infrastructure for institutional markets.
AI + CRYPTO
Crypto After the Speculation Cycle: The AI-Driven Institutionalisation of Digital Assets
Artificial intelligence will accelerate the institutionalization of digital assets by improving compliance, surveillance, custody risk management and market intelligence.
AI + CRYPTO
Stablecoins, AI Agents and the Future of Machine Payments
How stablecoins and AI agents could create the early infrastructure for machine-to-machine commerce.
AI + HEDGE FUNDS
The AI Hedge Fund Stack: From Alternative Data to Agentic Research
Hedge funds compete on speed, information asymmetry and judgment. Artificial intelligence changes all three.
AI + HEDGE FUNDS
Alpha in the Age of Machines: Why Human Judgment Becomes More Valuable, Not Less
As analytical tools become more powerful and widely available, human judgment becomes a more valuable differentiator.
AI + HEALTHCARE
Clinical AI at Scale: Why Healthcare's Real Opportunity Is Workflow, Not Hype
Healthcare AI should be judged by safer workflows, reduced administrative burden, better clinical capacity and measurable patient value.
AI + HEALTHCARE
From Medical Devices to Medical Intelligence: The Governance Challenge
Governance, validation, post-market monitoring and accountability will determine whether adoption creates value or risk.
AI + HOSPITALITY
The AI Concierge Economy: Why Hotels Must Rebuild Around Prediction
Hospitality has always been built on service. AI does not change that. It changes how service is anticipated.
AI + HOSPITALITY
Revenue Management 3.0: AI as the New Operating System of Hospitality
AI can connect pricing, distribution, loyalty, staffing, guest preferences and operations into one intelligent commercial system.
AI + REAL ESTATE
The AI Asset Manager: Real Estate Moves From Static Valuation to Live Intelligence
AI will shift institutional real estate toward live intelligence across assets, tenants, operations and capital allocation.
AI + REAL ESTATE
AI Infrastructure and the New Geography of Real Estate Value
AI is a physical infrastructure cycle: data centers, power access, cooling, land and connectivity are reshaping real estate value.
AI + GOVERNMENT
AI Government: From Digital Services to Predictive Public Administration
Digital government moved services online. AI government can make public administration more predictive, adaptive and responsive.
AI + GOVERNMENT
Sovereign AI: The New Infrastructure of State Capacity
Sovereign AI is becoming a question of state capacity, competitiveness, cybersecurity, public trust and strategic autonomy.
Research collection · RWA tokenization
Real-world asset tokenization, layer by layer
Researching how legally enforceable rights in real-world assets can be represented, administered and transferred through programmable digital infrastructure. Institutional research across real estate, private credit, hospitality, luxury assets, art and infrastructure, covering legal structures, financial engineering, token taxonomy and risk frameworks.
Institutional tokenization architecture
Each layer must be independently valid and collectively coherent. A failure in any single layer can compromise the entire structure.
Real-world asset tokenization is the coordinated representation of identifiable economic or legal rights through a digital token and an enforceable off-chain structure. Tokenization is not merely converting an asset into a cryptocurrency.
A token does not automatically create legal ownership of the referenced asset. Enforceability depends on the governing legal structure, issuance documentation, register of ownership, custody model and applicable law.
The strategic significance of tokenization is not the digital representation of an asset in isolation. It is the possibility of rebuilding issuance, ownership administration, compliance, settlement and asset servicing as a coordinated programmable system.
Tokenization lifecycle: thirteen operational stages from asset identification to redemption
- 01Asset Identification
- 02Rights and Cash-Flow Mapping
- 03Legal Feasibility
- 04Issuer or SPV Structuring
- 05Valuation and Due Diligence
- 06Token Classification
- 07Smart-Contract Architecture
- 08Compliance and Investor Eligibility
- 09Primary Issuance
- 10Custody and Settlement
- 11Asset Servicing and Reporting
- 12Transfers or Secondary Market
- 13Redemption, Maturity or Enforcement
AI-powered risk analysis covering crypto fraud, financial crime, compliance failures, cybersecurity threats, geopolitical risk and institutional due diligence. Evidence-based intelligence for decision-makers.
Research methodology
Quantitative analysis, institutional review, strategic foresight
Our research framework combines quantitative analysis, institutional review, and strategic foresight to deliver actionable intelligence.
Each research pillar is selected for its potential impact on enterprise architecture, institutional decision-making, regulated industries or long-term technological transformation.
The research is designed to connect AI concepts with real operating models, workflows, governance structures and executive decision systems.
Every pillar incorporates governance, accountability, auditability and risk considerations as core components rather than afterthoughts.
The research takes a multi-year perspective, examining how AI will reshape industries, institutions and economic structures over the coming decade.
Investor access
This section is reserved for qualified, professional or otherwise eligible investors. Enter your access code to open the NEO Protected SAFE documentation.
Code not recognised. Check it and try again. A secure connection (HTTPS) is required to open the investor area.Strictly confidential. Not a public offer. All other pages of this site remain public.