FLS26 Hackathon
2nd Placepacemaker.ai — AI-driven decision support for commodity traders
The Challenge
The pacemaker.ai challenge asked teams to design and prototype a solution that helps commodity traders, procurement teams, or market analysts make better decisions under uncertainty. The goal: turn raw market and news signals into actionable decision support around questions like buy now vs. buy later, hedge vs. stay exposed, stock up vs. stay lean.
We were given 5–6 years of LME metal price data and GDELT commodity news data. Everything else — architecture, output format, tools — was open.
Our Solution
We built a decision cockpit focused entirely on high-stakes trading decisions. The emphasis was on explainability and actionability — not just charts.
- —Focused UI — decisions only. Buy / wait / hedge. No noise, no dashboards for the sake of dashboards.
- —Explainability first — logistic regression, XGBoost, and decision trees so traders understand the "why" behind every recommendation, not just a signal.
- —Data edge — we integrated a public metal production plant thermal index dataset that wasn't in the provided data, giving a meaningful signal advantage.
- —Clean architecture — offline ML pipeline generates a snapshot; a FastAPI layer serves it; Streamlit renders it. Deployed on Cloudflare Pages. No live inference during requests.
Stack
| Layer | Technology |
|---|---|
| ML models | Logistic Regression, XGBoost, Decision Trees |
| Data | LME prices, GDELT news, thermal index dataset |
| Backend API | FastAPI (uvicorn) |
| Frontend | Streamlit |
| Containerisation | Docker |
| Deployment | Cloudflare Pages |
| Language | Python |
Architecture
The API is a thin backend-for-frontend layer — it never trains or runs inference during a request. Offline batch scripts compute price features, news sentiment scores, model predictions, and buy/hold/hedge actions, then write everything into a single artifacts/dashboard_snapshot.json. The API reads that snapshot and serves frontend-friendly JSON.
GET /api/dashboard GET /api/metals GET /api/metals/{metal}/recommendation GET /api/metals/{metal}/explanation GET /api/metals/{metal}/news The Team
It's rare to find such instant synergy with a new group of people. I'm grateful to have collaborated with Arda Rênas Ayik, Anton, Alexander Hülsmann, Akshit Bhatia, and Karamoko Assamoi — we built something real under real pressure and genuinely enjoyed every moment of it.
Huge thanks to Emilio Boldt, Carolin Schäfer, and the CHEFTREFF team for organising everything — from team building and tool access (Lovable, OpenAI, ElevenLabs, Langdock, Miro) to the speed dating with companies. CHEFTREFF has enormously scaled the opportunities for young talent and it shows.