DammSmartBuy is a procurement intelligence platform built for Damm's buying team. It answers one question: should we buy, wait, hedge, or monitor each commodity right now — and why?
What inspired us? Procurement decisions at a brewery are high-stakes and time-sensitive. Barley, aluminium, PET packaging, and energy together represent a significant share of production costs. We wanted to build something a real procurement manager could open on Monday morning and immediately act on — not just a dashboard full of charts, but a tool that gives a defensible recommendation with full source traceability.
How we built it We combined three layers: 1. Signal engine — each commodity has a weighted scoring model (0–100) built from documented market signals sourced through Cala.ai MCP knowledge search: ING Commodities Outlook, ICIS, OMIP, USDA, CFTC COT data, Strategie Grains, JRC MARS, Copernicus C3S, and more. High score = strong upward pressure = hedge. 2. ML forecasting — a Random Forest + Linear Trend ensemble trained on real price data. Barley uses 20 years of Damm's internal weekly data (2006–2025). Aluminium, Energy, and PET use FRED public datasets. The model is recursive: each predicted week is added to the training set before predicting the next. 3. Explainability layer — every recommendation comes with signal drivers (what's pushing prices up or down), a price vs EU inflation comparison, historical pattern matching (Pearson correlation of the last 8 weeks against all historical windows), and for Barley, a forward-looking drought risk monitor for the top producing countries. The interface is a 3-level Streamlit dashboard: category bubbles → sub-items → full commodity analysis with AI resume, KPIs, charts, and pattern matching.
What we learned Cala.ai MCP is a genuinely powerful tool for structured market intelligence — it let us source and verify dozens of signals in hours rather than days. Recursive ML forecasting is much more honest than single-shot prediction for procurement use cases. The hardest part of a decision-support tool is not the model — it's making the output actionable and explainable for a non-technical user.
Challenges No public weekly price data exists for PET — we used Brent Crude as a feedstock proxy and labelled it clearly. Aluminium and Energy public data (FRED) is monthly, so we interpolated to weekly before training. This reduces granularity but preserves structural trends. Balancing technical depth with UI simplicity for a business user audience.
Built With
- api
- cala.ai
- forest
- fred
- mcp
- pandas
- plotly
- python
- random
- scikit-learn
- streamlit
Log in or sign up for Devpost to join the conversation.