AI-Powered Bid & Proposal Response Engine

A working prototype for CUST Hackathon 2026 — Procurement, Sourcing & Contract Management track.

Ingests an RFP/RFQ/Tender (PDF/DOCX), extracts mandatory requirements, deadlines, budget and evaluation criteria, matches them against a Company Capability Library via RAG, drafts a structured proposal response, flags compliance gaps, and scores the opportunity's win-probability with a GO / NO-GO recommendation.

Quick start

cd backend
pip install -r requirements.txt --break-system-packages   # if needed
uvicorn app:app --reload --port 8000

Open http://localhost:8000 in a browser. A demo workspace ("PITB ICMS - District Case Management") is already loaded with the included sample RFP so you can explore the full pipeline immediately -- or click New RFP workspace and upload your own PDF/DOCX.

Optional: enable LLM enrichment

Set ANTHROPIC_API_KEY before starting the server to turn on:

• LLM-based cleanup / labeling of extracted requirements
• LLM-polished proposal draft prose

export ANTHROPIC_API_KEY=sk-ant-...
uvicorn app:app --reload --port 8000

Without a key, everything still works end-to-end using the deterministic rule-based extraction, TF-IDF RAG matching, and template-based drafting -- the demo never breaks due to missing connectivity.

Architecture

sample RFP (PDF/DOCX)
        |
        v
 extraction.py   --  regex/NER-lite: deadlines, budget, mandatory
        |             requirements, evaluation criteria, Q&A sections
        |             (+ optional Claude enrichment)
        v
   rag.py        --  TF-IDF retrieval over capability_library.json
        |             -> compliance checklist (pass / partial / gap)
        v
 analysis.py     --  derives win-probability model features
        |             (cert match %, requirement coverage,
        |              budget alignment, competitor count, ...)
        v
 scoring.py      --  logistic regression trained on historical_bids.csv
        |             (120 past bids) -> win probability + GO/NO-GO
        v
 proposal.py     --  drafts each Q&A section using matched evidence
        |             (+ optional Claude polishing)
        v
   export.py     --  assembles everything into a downloadable .docx
        |
        v
   app.py (FastAPI) + frontend/ (vanilla HTML/CSS/JS SPA)

Data (backend/data/)

Sample RFP

sample_rfps/PITB_ICMS_RFP_2026.docx -- a realistic IT-services tender (17 mandatory requirements, 8 weighted evaluation criteria, 5 Q&A sections, deadlines and budget figures) used to demo the full pipeline.

Mapping to deliverables

• Working prototype accepting a sample RFP -> structured draft response -- proposal.py + Draft tab
• Separate workspace per RFP/RFQ/Tender -- storage.py (one JSON record per workspace)
• Auto-generated compliance checklist with pass/fail mapping -- rag.py -> Checklist tab
• Win-probability dashboard -- scoring.py -> Dashboard tab, with explainable per-feature contributions
• GO / NO-GO decision -- scoring.go_no_go_decision()
• Review/edit/approve UI before export -- editable draft textareas + "Mark approved" + .docx export

Mapping to rubric

• Technical implementation -- full pipeline runs end-to-end offline (no external dependency required for the core demo); LLM enrichment is additive, not load-bearing.
• Validation & reliability -- the scoring model reports its own training accuracy (87.5%) and base win rate (63.3%) directly in the UI; the compliance checklist shows similarity scores for every match.
• Innovation -- combines rule-based NER, TF-IDF RAG, and an explainable logistic-regression scoring model with optional LLM enrichment -- degrades gracefully at every layer.
• Feasibility -- entirely Python/FastAPI + vanilla JS, runs on a laptop with no GPU and no mandatory API key.

Built With

• css
• html
• javascript
• python