A real-world discovery and coordination product built on a privacy runtime and a learning system.
AVRAI
privacy-first discovery and model infrastructure
Privacy-first discovery and coordination
AVRAI is building a system for discovering places, coordinating plans, and learning what works in the real world.
AVRAI combines three prongs: product applications, a control plane for identity and consent, and a compact reality model that improves ranking, planning, and coordination from observed outcomes.
Designed for local-first inference, bounded cloud use, and evidence-gated releases.
What AVRAI is
A product stack, a control stack, and a learning stack.
Applications deliver product value. Runtime governs trust and release. The reality model improves decision quality.
User product
AVRAI helps people discover places, build lists, coordinate plans, and follow through in the real world.
Operator product
The same stack supports reservations, host tooling, venue operations, and local workflow intelligence.
Privacy runtime
Identity, consent, transport, and rollout stay in a control plane instead of being scattered across app code.
Learning system
The reality model improves ranking and planning from saves, visits, returns, and attendance outcomes.
Three prongs
Separate the jobs. Connect them through one contract boundary.
Applications own experience. Runtime owns control. The reality model owns decision quality. The system only scales if those responsibilities stay distinct.
System map
three prongs / one contract boundary
Shared contractsData boundary and release boundary across all three prongs
Selected prong
Reality model
click a prong to inspect its role
Makes the stack learn from outcomes.
This prong improves ranking, forecasting, and next-action quality from lived outcomes.
State representation, scoring, transition prediction, and planning.
Behavioral signals, contextual state, and runtime-approved inference windows.
Scores, forecasts, and model candidates that can outperform heuristics.
Tech stack
A product stack with narrow boundaries.
Product surfaces sit at the top. Control stays in the runtime. Decision quality sits in the model layer. Cloud services stay narrow.
Product surfaces
Consumer and operator applications
- Discovery, lists, shared planning, reservations, and local workflow tools
- Shared contracts across mobile products, operator tools, and the public web
Control plane
Identity, policy, transport, and release
- Consent, authorization, encrypted transport, rollout, rollback, and recovery
- Inference, sync, and operator actions execute through runtime gates
Decision models
Compact models for ranking and planning
- State representation, scoring, transition prediction, and planning
- Local-first inference with bounded training, evaluation, and promotion
Security + cloud
Narrow cloud responsibilities
- Model delivery, encrypted sync, observability, and privacy-preserving aggregation
- Security services and key management for local-first and distributed operation
Reality model
The reality model is AVRAI's learning engine.
This is the layer that lets AVRAI move past static retrieval and hand-tuned heuristics. It learns which places fit, which plans are likely to hold, which reservations convert, and which actions improve follow-through.
Why it matters
It turns outcomes into better product decisions
The model improves ranking, planning, and coordination using observed behavior instead of fixed heuristics alone.
What it learns
Fit, transitions, and follow-through
It models people, places, groups, context, and likely next states to estimate relevance, risk, and expected completion.
Why this design
Built for bounded operational decisions
General-purpose LLMs optimize for broad generation and reasoning. World models optimize for environment simulation. AVRAI optimizes for low-latency decision quality in one operating domain.
Observe outcomes
Collect signals such as saves, visits, dismissals, returns, reservations, and attendance.
Build compact state
Encode users, entities, and contexts into a representation the model can score and simulate.
Train and compare
Learn scoring and transition behavior, then shadow it against incumbent heuristics before promotion.
Ship under gates
Promote only the models that clear privacy, rollout, rollback, and measurable outcome requirements.
That is the compounding advantage: AVRAI can improve from lived outcomes while keeping privacy, latency, and release control intact.
Early access
Join the AVRAI waitlist.
Priorities: product pilots, operator design partners, and research conversations.