VKD/Start a conversation
Back to projectsINTERACTIVE PRODUCT DEMO / REPRESENTATIVE DATA

FOODFLOW / DECISION ENGINE

Waste signals become
operational decisions.

Explore the serving layer of a shelf-life-aware recommendation and anomaly-diagnostics system.

View executable source
01
Recommendation requestApply operational constraints before serving
RESULT
4 constrained recommendationsPerishability rule active
01
Organic Whole Milk7 day shelf life · Medium perishability
14.0
02
Cheddar Cheese Block21 day shelf life · Low perishability
9.0
03
Greek Yogurt12 day shelf life · Medium perishability
8.4
04
Rolled Oats180 day shelf life · Low perishability
6.5
02
Anomaly diagnosticGold-layer observation versus operational baseline
ANOMALY DETECTED+37.8%

Cold-chain variation or replenishment mismatch

Recommended action

Inspect refrigeration sensors and reduce the next order by 15%.

ENGINEERING EVIDENCE

What the full implementation proves.

<15 msReported FastAPI response latency
3 layersBronze, silver and gold data model
0% leakageShort shelf-life items under suppression tests
IoT + ordersRaw telemetry and transactions
→
DuckDB lakehousedbt bronze → silver → gold
→
Decision servicesRecommendations, diagnostics, nudges
→
API consumersOperations and inventory systems

AUTOMATED VALIDATION

Quality gates

  • Perishability-constraint assertions
  • Cold-start fallback for unseen users
  • FastAPI application compilation
  • Docker build in GitHub Actions

PRODUCTION TRADEOFFS

Decisions at scale

DuckDB → cloud warehouse

Embedded analytics keeps the project portable; multi-team production workloads would move durable facts to Snowflake, BigQuery or Redshift.

Batch scoring → feature service

Precomputed affinity is predictable and inexpensive; high-velocity personalization would require streaming features and online model monitoring.

Local rules → governed policy

Shelf-life constraints are explicit and testable; enterprise deployment needs versioned policy ownership, audit logs and controlled overrides.