ProposalIQ

About this project

• •Federal solicitations turn on exact identifiers — clause numbers, CLINs, section references. Pure semantic search kept missing them, which is the reason this system is hybrid rather than dense-only.
• •A production RAG pipeline over Qdrant fuses dense embeddings (text-embedding-3-large, 3072 dimensions) with BM25 sparse retrieval using Reciprocal Rank Fusion at k=60. I chose hybrid after watching semantic search repeatedly miss the exact identifiers the documents hinge on.
• •Claude, GPT-4o and Gemini sit behind a unified layer with per-model token and cost accounting. Opportunity sourcing is agentic: SAM.gov via MCP tools driven by a Claude tool-use loop, alongside the HigherGov REST feed.
• •I built an evaluation subsystem for A/B model judgments and weighted rubric scoring across writing quality, completeness and accuracy, surfacing win-rate and tie analytics. Prompt and model changes get compared on results instead of intuition.
• •React SPA, an Express API with 162 endpoints across 33 data models, and a Python retrieval service. SSE-streamed drafting, real-time collaborative DOCX editing over Socket.IO, 442 automated tests, and CI/CD to Oracle Cloud through GitHub Actions, pm2, Nginx and a Dockerized OnlyOffice.

Key Features

• •Dense and sparse retrieval, fused with Reciprocal Rank Fusion
• •Three model providers behind one accounted layer
• •Agentic opportunity sourcing over SAM.gov via MCP
• •Model choices decided by measurement, not intuition
• •Real-time collaborative DOCX editing over Socket.IO
• •442 automated tests and CI/CD to Oracle Cloud

Technologies Used