StormOps

Autonomous factory-ops agent for weather-driven supply-chain disruption.

On a trigger, StormOps pulls a real weather forecast for supplier regions, reasons over a supplier/shipment graph to score disruption risk, auto-executes low-risk mitigations, escalates high-risk ones to a human, and fires the approved actions as real Slack/email — every step logged. Built for the Harness Engineering Hack.

It also maintains a cached supply-chain weather monitor: GET /supply-chain/weather uses Open-Meteo, optional WeatherAPI, wttr.in, and free search context for country and route-level disruption snippets, refreshing every 4 hours by default. Uploaded supply-chain CSVs become the active graph Claude can inspect through route-weather tools.

One-liner: StormOps turns real-time extreme-weather signals into governed, human-approved supply-chain mitigation actions across procurement, logistics, and factory ops.

---

Why this wins (maps to all 5 judging criteria, 20% each)

• Idea — real factory/procurement pain, not another chatbot.
• Technical — real weather data + a multi-agent reasoning pipeline.
• Tool Use — coherent story across Jua, Composio, OpenUI, Guild, Render.
• Presentation — one clean 3-min storm scenario.
• Autonomy — runs on a trigger over real-time data; tiered autonomy auto-acts on low-risk items and only escalates expensive ones. This is the criterion most teams fumble — lean on it hard.

Architecture

NL/trigger event
   └─ Weather Agent   (Jua real forecast → wind/precip/temp/severity)
        └─ Impact Agent   (severity × criticality × urgency → at-risk shipments)
             └─ Mitigation Agent   (reroute / expedite / buffer)
                  ├─ low-risk  → auto-execute → Comms Agent (Composio: Slack/email)
                  └─ high-risk → human approval → Comms Agent
   all governed + triggered by Guild ; every event → ClickHouse ; UI via OpenUI ; deploy on Render

Run (mock mode, zero keys)

cd backend && pip install -r requirements.txt
python main.py            # no-server smoke test, prints a full incident
uvicorn main:app --reload # API: POST /run {"event": "..."}, POST /approve, GET /events

Supply-chain weather:

curl "http://localhost:8000/supply-chain/weather"
curl "http://localhost:8000/supply-chain/weather?force_refresh=true"
curl "http://localhost:8000/supply-chain/weather/routes?supplier_id=S1"

---

Demo-safe build order (do them in this order; you're shippable at every checkpoint)

1. 0:00–0:30 — Core loop (DONE in this repo). python main.py already produces an incident with auto-executed + pending actions. You have a demo now.
2. 0:30–1:15 — Composio real Slack send. Wire the block in comms_agent. Grab a key from the Composio DevRel. First real action = locks the Autonomy story + the $200.
3. 1:15–2:00 — OpenUI adaptive front-end. npx skills add thesysdev/openui then have Claude Code scaffold an incident console that reads /run + /events. Make it adaptive: calm card for moderate, full red console for severe. ($2,000, few use it.)
4. 2:00–2:45 — Jua real weather. pip install jua, key from Jua DevRel/CEO. Replace the mock in weather_agent. Now "acts on real-time data" is literally true. (Judge = CEO.)
5. 2:45–3:45 — Guild wrap + trigger. Register the deployed app in Guild, run it on a schedule trigger, route pending_approval through Guild's HITL gate. ($2,800 headline.) If the SDK fights you, the web-UI + Render-deploy + schedule path still earns the prize.
6. 3:45–4:15 — Render deploy. Push backend + frontend. (Free credits, easy mention.)
7. 4:15–4:30 — ClickHouse (if green). Swap log_event to insert into agent_events, show the timeline in the UI. ($1,600, judges love observability.)
8. Stretch — Airbyte. Replace ONE mock pull with a real connector (e.g. procurement backlog from Linear/Jira). Only if everything above is solid. ($1,750.)

Cut order if behind: drop Airbyte, then ClickHouse, then Guild-SDK (keep Guild web-UI). Never cut: Composio (real action) + the adaptive UI + the autonomy framing.

The 3-minute demo script

1. Type: "A massive storm front is hitting Central Europe."
2. Weather Agent (Jua): Wind 92 km/h, Rain 47 mm, severity HIGH — real forecast.
3. Impact Agent: 3 suppliers affected, 3 shipments at risk, ranked by score.
4. Mitigation Agent: proposes reroute/expedite per shipment.
5. Autonomy beat: the $80k reroute auto-executes (Slack fires live via Composio); the $120k and $150k ones pause for human approval.
6. Click Approve on one → real Slack message sends.
7. Show the Guild trace + ClickHouse timeline: incident resolved, every action audited. Close: "Jua sees the storm, the agent reasons about your supply chain, Composio acts, Guild governs, OpenUI adapts, Render runs it — autonomously, with a human gate on the expensive calls."

Keys to collect on-site (the DevRels are the speakers/judges)

• Composio — Slack/Gmail send
• Jua — weather API (X-API-Key)
• OpenUI/Thesys — C1 key if using the hosted GenUI engine
• Guild — workspace + (Python or TS) SDK access
• ClickHouse / Render — accounts