Operational memory for recall containment. Turn a real FDA recall into a scoped blast radius, a Gemini-drafted action plan, human approvals, and an audit-ready compliance report — with a tamper-evident audit trail.
Google Cloud Rapid Agent Hackathon · Fivetran track · MIT licensed.
🔗 Live demo: https://recall-ops.vercel.app · Repo: https://github.com/vaibhav4046/RecallOps
📘 Project whitepaper: docs/RecallOps-Cortex-Whitepaper.pdf · a full walkthrough of the goals, the use cases stated brutally, the architecture, the data model, the agent design, the audit chain, and operations (LaTeX source: docs/RecallOps-Cortex-Whitepaper.tex).
The hosted URL serves the frontend in clearly-labelled demo / fallback mode; it flips fully live (real openFDA recall + live Gemini) once the backend API is deployed to Cloud Run and
window.RO_CONFIG.apiBaseinindex.htmlpoints at it.
RecallOps Cortex pulls a live openFDA food-enforcement recall, scopes it across operational data (stores, inventory, POS, shipments, customers) synced through Fivetran into BigQuery, has Gemini 3 draft containment actions, gates every external action behind human approval, and writes every decision to an append-only sha256 hash-chain audit log.
It is not a chatbot and not a static mockup — it is a real backend (FastAPI on Cloud Run) behind a command-center UI. The recall is always live from openFDA; cloud integrations activate with credentials and otherwise run in clearly-labelled fallback, so the demo never breaks.
flowchart LR
FDA["openFDA Food Enforcement API"] -->|live recall| API["Cloud Run API · FastAPI"]
OPS["Seeded ops systems<br/>POS · inventory · shipments · stores · customers"] --> FIV["Fivetran MCP · 161 ops"]
FIV --> BQ[("BigQuery<br/>operational warehouse")]
API <--> BQ
API <-->|reason · drafts only| GEM["Gemini 3 · Agent Builder / ADK"]
API --> OTEL["OpenTelemetry"] --> OBS["Arize Phoenix · Dynatrace"]
API --> UI["RecallOps Cortex frontend"]
UI --> HUM{"Human approval gate"}
HUM -->|approve / reject| API
API --> AUD[("Audit hash-chain")]
API --> REP["Compliance report"]
Agent flow: intake (real openFDA recall) → sync (Fivetran → BigQuery, freshness-gated)
→ scope (blast radius across the warehouse) → reason (Gemini drafts action cards, cites
evidence) → review (human approves each, no auto-execution) → report (audit-ready).
| Table | Key columns |
|---|---|
recalls_raw |
recall_id · classification · product_description · recalling_firm · reason · distribution_pattern |
inventory_lots |
location_id · sku · lot · units |
pos_sales |
customer_id · sku · lot · units · consented · channel |
shipments |
id · supplier_id · location_id · sku · lot · status · quantity |
store_locations |
id · name · type · city · manager · risk_tier · lot · sku · units |
containment_actions |
action_id · type · priority · owner · scope · evidence_ids · approval_state · status |
audit_events |
event_id · actor_type · actor_name · event_type · label · evidence_ref · timestamp · prev_hash · hash |
agent_runs |
run_id · model · prompt_version · tool_count · latency_ms · token_count · eval_score · status |
Blast-radius runs as real BigQuery jobs over inventory_lots / pos_sales / shipments / store_locations;
their job IDs are surfaced as proof. The audit_events hash chains each event to the previous one
(hash = sha256(prev_hash + event)), so any tampering breaks the chain.
- Real recall —
GET /api/recalls/latestpulls a genuine openFDA Class I record (no key). The UI shows the live record id + retrieval time. - Real blast radius — aggregated over seeded rows bound to the recall's lot codes, with live sync/query timestamps (and real BigQuery job IDs in live mode).
- Human-approval gate — nothing executes until a human approves; each approval
writes a real audit event. End-to-end verified: 6 approvals →
SYSTEM: CONTAINED→ audit chainintact: true. - Honest status — every integration is badged Live or Fallback — no green light unless it's actually wired.
- Real observability — the agent pipeline emits OpenTelemetry spans, viewable
in
/llmopsand exportable to Arize Phoenix / Dynatrace.
Every integration is LIVE or PLUGGABLE — never faked. The /architecture
page renders this from real backend config.
| Partner | Role | Status |
|---|---|---|
| Google Cloud | Gemini 3 · Agent Builder / ADK · Cloud Run · BigQuery | core |
| Fivetran | MCP sync (161 ops) → BigQuery | live / pluggable |
| Arize Phoenix | LLM tracing + evals · OpenTelemetry | live (OSS, local) |
| Dynatrace | APM / observability via OTLP export | otlp-ready |
| Elastic | Recall + audit full-text search | pluggable adapter |
| MongoDB | Agent memory + vector similar-recalls | pluggable adapter |
| GitLab | CI/CD + DevSecOps pipeline (.gitlab-ci.yml) |
in repo |
RecallGraph, a 3D operational-memory graph expanding one FDA recall into SKUs, lots, stores, shipments, customers, actions, evidence, memory, evals, and improvement proposals.
# backend API (:8099) — real openFDA, fallback for the rest
python -m venv .venv
.venv/Scripts/python -m pip install -r backend/requirements.txt
.venv/Scripts/python -m uvicorn app.main:app --app-dir backend --port 8099
# frontend (:8790), second terminal
python -m http.server 8790Open http://127.0.0.1:8790. Health: http://127.0.0.1:8099/api/health.
Smoke test the whole flow: python backend/scripts/smoke_test.py → RESULT: PASS.
Going live (Gemini 3 / BigQuery / Fivetran / Cloud Run deploy): see
docs/setup.md — copy-paste commands, ~$0 on free trials.
GET /api/health · GET /api/recalls/latest · POST /api/recalls/select ·
POST /api/fivetran/sync · GET /api/fivetran/status · POST /api/blast-radius ·
POST /api/agent/run · GET /api/agent/runs · POST /api/actions/:id/approve ·
POST /api/actions/:id/reject · GET /api/audit/:recallId · POST /api/report/:recallId ·
GET /api/ecosystem · GET /api/state. Every response carries a trace_id.
/ command center · /radar recall radar · /fivetran control tower · /graph 3D
RecallGraph · /context context pack · /evidence evidence board · /actions approval
workbench · /llmops LLMOps tower · /improvement self-improvement · /replay crisis
replay · /compliance audit timeline · /report compliance report · /architecture ·
/settings.
- No external action without human approval — enforced server-side.
- Tamper-evident audit — every step timestamped with actor + evidence; the sha256 hash chain breaks if any event is altered.
- No fake claims — cached/fallback data is labelled; only real integrations show "Live".
cortex/
index.html · theme.css · *.jsx frontend (React via CDN, no build step)
lib/data.js · lib/api.js seed + live API client (with labelled fallback)
backend/app/ FastAPI: openfda · warehouse · gemini · bigquery ·
fivetran · elastic · mongodb · telemetry · audit · state
backend/scripts/ smoke_test · sanity_check · load_bigquery
.gitlab-ci.yml lint → test (+SAST) → human-gated Cloud Run deploy
docs/ architecture · setup runbook · submission pack
| Command center | RecallGraph (3D) |
|---|---|
 |
 |
| Agent reasoning + actions | Compliance / audit |
 |
 |
MIT — see LICENSE.