EvidenceDeskAI Support Copilot with Human Review
Independent Developer
I independently designed and built EvidenceDesk, an AI support copilot that turns a support ticket into a source-backed draft note for human review. The application connects document retrieval, structured AI responses and an explicit approval workflow in a complete web product.
Read full description: EvidenceDesk | AI Support Copilot with Human Review
I developed the Python backend with FastAPI and a bounded LangGraph workflow, combining PostgreSQL full-text search with pgvector similarity search and reciprocal rank fusion. Cloudflare Workers AI provides embeddings and generation, while citation and quotation checks connect generated claims to the retrieved documentation.
I built the Next.js and TypeScript interface for ticket analysis, source inspection, approval or rejection, and saved-note history. Session and tenant authorization restrict access to records. Approval saves the exact reviewed proposal in a database transaction, with retry handling designed to prevent duplicate notes; model output cannot directly execute a write.
The project includes automated API, database and browser tests, CI/CD, and a published evaluation record that separates retrieval results from answer-quality assessment. EvidenceDesk is an independent portfolio demonstration using fictional support cases. Its evaluation documents limitations, and human review of semantic answer quality remains pending.



