Delectable AI · Intelligence & ML

The grocery intelligence engine behind the agents — model portfolio, governed runtime, data substrate & learning loop.
Recommenders are composed score functions, grounded by the graph — the LLM is the last-mile narrator, not the brain. Every model call is authorized, budgeted & metered by the control plane; governance is the authority. We win on breadth × depth × moment-precision, not model size or data volume.
GroundedGovernedComposedModel-agnostic
Composed, not a black box
Candidate generator → ranker → policy filter → business objective → evidence. The LLM narrates the result; it never decides unchecked.
Control plane · spans every model
🛡️ Model Governance & Control Plane
The Policy Decision Point every model call passes through (ml control plane :8109 → governance :8095) — no model runs unauthorized, unbudgeted, or unmetered, and every recommendation is replayable.
Model entitlementBudget & meteringContent gate (deny-wins)Model provenanceEvaluation authorityAudit & explainabilityDomain packs
Roadmap & Vision — widening the intelligence moat
In production
Recommender orchestration (model-of-models) · Lift-over-baseline propensity + precedence fusion · Grounded flavor-pairing engine · Perfect-Cart ILP · Multi-stage ranking · BQML mission clustering · Graph analytics (PageRank · Leiden · Node2Vec) · Grounded RAG · Governed model registry + entitlement · Custom-SLM fine-tuning pipeline · Holdout incrementality.
In build
GNN link-prediction (substitution / pairing) · Category / edge expert SLMs · Session / real-time propensity + online feature store · Aroma-compound data (CoSyLab / FlavorDB) for molecular flavor discovery · Household persona confidence · CV produce-quality & label vision · Survival-model replenishment · Learning-to-rank · Uplift modeling.
12–24 months
Distill frontier reasoning into compact category SLMs · Causal retail-media incrementality + RL bidding (true ROAS) · Multi-objective optimization (click + convert + retain + margin) · Contrastive cold-start (new SKUs / tenants) · DPO from audited corrections · Federated, privacy-preserving cross-tenant learning · Health-outcome & clinical-diet personalization.
↗ = planned / not yet in production.
Model types in the portfolio
llm-embeddinggraph (Node2Vec · MAGE)bqml (K-means · Bayesian)statistical (PMI)optimizer (ILP · CP-SAT ↗)rule-based scorerllm function-callingllm-as-judgegrounded RAGcustom SLM ↗GNN ↗
The thesis: intelligence is a portfolio of grounded, governed models the agents compose — the moat is fusing many dimensions and activating the right ones for this shopper at this moment, not any single model. Internal source-of-truth inventory: doc/ml-model-catalog.md; governed registry: ml/domains/<domain>/model-manifest.yaml.