Capabilities behind the agentic grocery experience.

01

Real-time propensity personalization

Behavioral and declared signals update the shopper profile during the session, so search, ranking, offers, and agent copy respond to what the shopper is doing now.

declaredsessioninferred

InputSearches, clicks, offer clips, cart events, dietary declarations, purchase history.

DecisionFuse hard preferences with live propensity changes and mission context.

OutputRanking weights, preference evidence, and personalized agent response context.

02

GNN householding and de-averaging

A loyalty ID often represents multiple eaters. Graph-based clustering separates overlapping needs into household personas instead of averaging them into one unusable profile.

graphpersonahousehold

InputBasket co-occurrence, recipe saves, category affinity, dietary conflicts, user events.

DecisionFind connected purchase clusters and split mixed households into targetable subgroups.

OutputPersona-aware suggestions, safer dietary handling, and better retail-media targeting.

03

Recommender orchestrator

The platform runs specialized models for healthier swaps, replenishment, meal kits, promotion fit, pantry completion, and mission-specific ranking, then selects the right mix by context.

15 modelsroutingranking

InputSurface, mission, profile, catalog candidates, business rules, historical outcomes.

DecisionScore each model's expected usefulness for the current shopper and task.

OutputA blended, explainable recommendation set with model-level provenance.

04

Virtual pantry decay forecasting

Purchase cadence, household size, recipe intent, and depletion curves predict what is still at home, what is low, and what should be replenished before the shopper asks.

forecastpantryreplenish

InputOrder history, package sizes, consumption estimates, explicit pantry edits, recipe activity.

DecisionPredict depletion and suppress unnecessary purchases while staging likely refills.

OutputPredicted cart items, skip reasons, stock status, and replenishment urgency.

05

Direct Preference Optimization feedback loop

Merchandiser corrections, rejected agent actions, compliance notes, and human preference edits can be converted into preference pairs and policy-aware training data.

feedbackbrand rulestraining

InputApprovals, rejections, edits, evaluator failures, brand guidance, policy flags.

DecisionTranslate observed corrections into structured preference evidence.

OutputTraining pairs, prompt updates, regression tests, and governance review trails.

06

Agentic evaluation and adversarial testing

Before an agent action reaches a shopper or operator, hard rules and judge models can verify safety, policy fit, evidence quality, and output structure.

evalred teamsafety

InputResponse candidates, action payloads, policy requirements, evidence, adversarial prompts.

DecisionRun deterministic guards first, then independent judges for semantic quality.

OutputPass/fail status, rubric scores, failure reason, and regression fixtures.

07

Annotate, do not blindly filter

When intent is ambiguous, the system can rank safe choices first while preserving commerce coverage with visible confidence annotations and shopper-context caveats.

confidenceambiguitytrust

InputDietary confidence, shopper intent, household ambiguity, product attributes, policy rules.

DecisionChoose when to suppress, rank down, annotate, or ask for clarification.

OutputReasoned labels, risk notes, alternate choices, and safer UI presentation.

08

Deterministic recipe-to-cart engine

Recipe text becomes a shoppable cart through parsing, canonicalization, category anchoring, pantry checks, product search, and package-size math before the LLM does any wording.

parsercartILP

InputIngredients, servings, household size, store catalog, package sizes, pantry state.

DecisionConsolidate units, match products, optimize package selection, and preserve traceability.

OutputSKU-backed cart lines, alternatives, price estimates, sizing method, and reasoning.