AI Bid and Proposal Engine
AI/MLAgentic RFP-to-proposal pipeline that drafts a proposal section by section and returns a win probability from historical bid outcomes.
- Year
- 2026
- Role
- Retrieval and win model
- Stack
- Python, XGBoost, RAG
What it does
Takes an RFP document, finds comparable past bids, drafts the proposal section by section, and returns a win probability alongside the evidence it used. Seven stages end to end. Built for a CUST hackathon problem.
My contribution
I owned the retrieval layer and the XGBoost win-probability model trained on historical bid outcomes.
The leak that looked like a perfect model
The win model came back with a validation AUC of 1.000. A perfect classifier almost always means the model is cheating and nobody has noticed. I traced the feature table and found a Score% column that only gets populated after a bid is awarded. The model was reading the answer.
I removed it, rebuilt the feature pipeline from pre-decision fields only, and the AUC dropped to something believable and properly calibrated. We shipped a worse number and a system you could actually trust. That trade is the point.