Applied AI / RAG
AI Expert Discovery Platform
Built a RAG-based expert discovery platform combining vector retrieval and LLM reasoning to move beyond keyword-only matching.
The Product Challenge
Finding the right expert through keyword search fails in predictable ways: people describe expertise in different words, and the best match is often someone who never used the exact search term.
The product needed to understand expertise semantically - retrieving and ranking people based on what their experience actually means, not the strings it contains.
The core challenge was not simply calling an LLM. Retrieval quality, vector search and reasoning had to work together as a product system.
My Engineering Contribution
- Built the production RAG pipeline: document ingestion, embeddings and vector retrieval on Qdrant.
- Developed Spring AI microservices and FastAPI services around the retrieval core.
- Combined vector search with OpenAI and Gemini reasoning to rank expertise beyond keyword matching.
- Owned the path from embedding pipelines through production deployment.
- Mentored teammates on RAG pipeline design.
System & Product Considerations
- Retrieval quality before model choice - chunking and embedding strategy set the ceiling
- Ranking that blends vector similarity with LLM reasoning
- Latency and cost per query as product constraints
- Evaluating relevance when there is no single correct answer
Technical Areas
What This Project Taught Me
- RAG quality is won or lost at retrieval. The LLM call is the easiest part of the system.
- Retrieval, ranking and reasoning have to be designed as one product system, not three separate components.
- AI features still need boring engineering: pipelines, monitoring and deployment discipline.