State
Encode user, entity, and context state
Build compact representations from locality, timing, lists, trust, and behavioral signals.
AVRAI
privacy-first discovery and model infrastructure
Model overview
The reality model is a compact pipeline for scoring and planning.
AVRAI separates representation, scoring, simulation, and action selection into a model stack that can run under real latency, memory, and privacy constraints. The point is better next-action quality through outcome-grounded learning.
Model pipeline
Four model jobs, kept legible.
AVRAI should not collapse representation, ranking, and planning into one opaque block.
Score
Use a learned energy surface
Replace many static formulas with one learned scoring function over state and candidate action.
Simulate
Predict how the state can change
Model taste drift, list evolution, attendance risk, and action consequences before promotion.
Act
Choose the next best sequence
Move from one-step ranking to planning across lists, groups, reservations, and operator actions.
Training loop
Improve the model without unbounded cloud collection.
Observe
Collect visits, saves, dismissals, returns, attendance, and operator outcomes.
Consolidate
Compress short-term traces into usable memory during low-friction windows.
Fit
Train state, scoring, and transition components under device and budget constraints.
Evaluate
Shadow new models against incumbent heuristics before anything is promoted.
Promote
Ship only the models that clear outcome, privacy, and rollback gates.
Evaluation path
How model changes earn release.
Offline
Representation and ranking tests
Check retrieval quality, ranking lift, calibration, and robustness before live exposure.
Shadow
Side-by-side model comparison
Run learned paths against incumbent heuristics to verify win rate and failure behavior.
Release
Evidence-gated promotion
Promotion requires outcome lift, drift review, privacy compliance, and rollback confidence.
Boundary discipline
See the privacy rules around the model.
The model is only meaningful if privacy, consent, and transport are enforced before promotion and sync.
PrivacyRelease horizon
See when model work moves from research into product.
The roadmap separates architecture, active model build, and longer horizon research.
Roadmap