Food Rescue Map

FoodRescueMapArchitecture
MainFlow
IntentTrackingFlow

Inspiration

Late markdowns become morning waste. Store teams know it, but they can’t keep up. In 2022, the world wasted about 1.05 billion tonnes of food—roughly one‑fifth of consumer‑available food (UNEP, Food Waste Index 2024). The issue isn’t awareness—it’s timing and trust. Manual markdowns are slow, inconsistent, and hard to audit. AI can help, but AI without traceability is a liability. Retailers need to show why a price changed, what data drove the decision, and who approved it. Food Rescue Map makes every AI pricing decision traceable to a Confluent Cloud Kafka event (topic / partition / offset), so decisions are fast, safe, and provable.

What It Does

MainFlow

Food Rescue Map turns live inventory streams into rescue offers through a fully auditable AI pipeline.

Core Flow

Inventory events stream through Confluent Cloud Kafka
AI agents generate safe markdown candidates
Human‑in‑the‑loop (HIL) approval prevents runaway pricing
Approved offers appear on the map in real time

Evidence‑First Design

Evidence is not a log. It’s a one‑click operational UI that traces: Input → Reasoning → Guardrails → Approval → Publish

Every decision displays its Kafka coordinates (topic / partition / offset), making the full chain auditable on real streaming data. Managers can filter to show only Kafka‑originated events, proving decisions flow from real inventory—not manual overrides.

Intent Tracking (Behavioral KPIs)

IntentTrackingFlow

Beyond offers, we track engagement signals through the same Kafka infrastructure:

Offer Views: which items attract attention
Map Opens: navigation intent (strong purchase signal)
Dwell Time: engagement depth per offer These metrics aggregate into per‑store conversion funnels via /api/ops/metrics.

Impact KPIs

Waste avoided (items rescued before expiry)
Margin uplift (revenue recovered vs. full discard)
Time‑to‑markdown (hours saved from manual process)

How We Built It

Confluent‑first streaming architecture on Google Cloud.

Architecture Overview

Architecture FoodRescueMapArchitecture

Component	Technology	Deployment	Purpose
Frontend	React (Vite + Tailwind) + Google Maps	Firebase Hosting	Map UI, Approvals, Evidence, Impact
Edge API	FastAPI	Cloud Run	REST endpoints, Kafka publishing, SSE
Stream Worker	Python + Kafka Consumer	Cloud Run	Event processing, AI orchestration
Messaging	Confluent Cloud Kafka	Managed	Event streaming, Schema Registry
State	Firestore	Managed	Projection (Kafka is source of truth)
Failure Isolation	DLQ topic	Managed	Dead letter queue for failed events
AI	Vertex AI (Gemini 2.5 Flash Lite)	Managed	Multi‑agent reasoning with strict JSON outputs

The 6‑Agent AI Pipeline

When inventory accumulates (item or time window), the Stream Worker auto‑triggers multi‑agent analysis:

Demand
Waste Risk
Strategy
Guardrail
Briefing
Policy Each agent outputs structured JSON validated against schemas, keeping downstream processing consistent and safe.

Intent Tracking Flow

Intent Tracking User interactions flow through the same Kafka infrastructure: User Action → Edge API → Kafka (frm.offers.events) → Stream Worker → Firestore event_log

Schema Registry Integration

Every event type is registered with Confluent Schema Registry (backward‑compatible versioning):

inventory.upload.v1
inventory.item.created.v1
offers.candidate.v1
offers.approved.v1
offers.viewed.v1
offers.map.opened.v1
offers.view.closed.v1

Challenges We Ran Into

1. Safe Automation for Real Pricing

AI‑driven pricing is high‑stakes. A runaway discount could destroy margins.

Solution: Two‑layer control:

AI Guardrails: hard limits on discount percentages
Human‑in‑the‑Loop: every recommendation requires approval

2. Guaranteed Traceability

We needed every AI decision traceable to source data.

Solution: Kafka offset metadata travels through the pipeline. Evidence displays kafkaOrigin: { topic, partition, offset, timestamp }.

3. Real‑time Updates Without Polling

Users expect instant updates when offers are approved.

Solution: Server‑Sent Events (SSE) from Edge API. Frontend subscribes to /events/stream.

4. Repeatable Demo Scenarios

Real‑time systems are unpredictable during demos.

Solution: Demo replay endpoints that replay scenarios through Kafka for reliable judging.

Accomplishments We’re Proud Of

Auto‑triggered by inventory events, with manual override for demos/ops
Full audit trail: every decision traceable to Kafka coordinates
Real‑time updates via SSE
Schema‑enforced contracts via Schema Registry
DLQ isolation for resilience

What We Learned

Event‑driven AI removes manual bottlenecks
Schema Registry + Kafka offsets create auditable AI
Responsible pricing needs guardrails + HIL
Single‑store value enables low‑risk pilots

What’s Next for Food Rescue Map

Pilot with 2–3 stores and A/B test AI recommendations vs baseline
Add store‑specific demand signals (seasonality, weekday, weather)
Expand real‑time intent signals (in‑app actions, map opens, dwell time)
Scale to multi‑store networks

Potential Impact (Hypothesis)

Target Customer: Urban bento/deli chains (30–50 items discarded per day per store) Hypothesis: AI + real‑time inventory monitoring can reduce waste by 20–30% Conservative Estimate: