Skip to content
← All projects

Case study · Internship project at VietUnion (Payoo) · SRS + full-stack

DDQ Knowledge Assistant

An on-premise RAG assistant that drafts due-diligence answers for analysts to review.

Internal system. This page describes the problem and approach at a high level only — no screenshots, data or detailed architecture.

Problem
Answering due-diligence questionnaires means searching many past documents by hand, and the source documents must stay inside the company.
Approach
  1. Wrote the SRS, then built the full stack: FastAPI backend, Next.js frontend, PostgreSQL with pgvector.
  2. Hybrid retrieval: embedding search and BM25 run in parallel, merged with Reciprocal Rank Fusion.
  3. Low-confidence questions are routed to an LLM to draft an answer. Users choose the drafting model: a local LLM via Ollama or the Gemini API.
  4. Every draft goes to a human for review. The AI suggests, the analyst decides.
Stack
FastAPINext.jsPostgreSQL + pgvectorSQL ServerOllamaGemini API
Outcome
Delivered as a working internal tool and handed over to the team when my internship ended. I left before it was evaluated on live questionnaires, so I don't report usage or accuracy numbers.