Portfolio RAG: Grounded Recruiter Q&A (Citations + Evals)

Recruiters and hiring managers want fast, trustworthy answers about experience and impact — but typical chatbots hallucinate. The goal was a grounded QA system that only answers using my portfolio + resume sources, with citations and measurable reliability.

• Ingested portfolio case studies + resume PDFs into a document pipeline (chunking + metadata).
• Embedded chunks into a vector index and added a retrieval layer (top-k + optional rerank).
• Generated answers with strict grounding: citations required; refuse when evidence is missing.
• Built an evaluation harness (golden Q/A) to score answer relevance + citation coverage + faithfulness.
• Instrumented usage telemetry to monitor latency, failure rates, and cost per query.

• Demonstrates production-style RAG thinking: grounding, evals, guardrails, and monitoring — not just a demo chatbot.
• Turns a portfolio into an interactive, verifiable knowledge base recruiters can trust.
• Shows AI PM skill set: system design tradeoffs (quality vs latency vs cost) and measurable reliability.

• Deployed as a web app with API routes for retrieval + generation.
• Guardrails: ‘answer only from sources’ + refusal fallback + max context limits.
• Monitoring: latency, cost/query, top queries, retrieval hit-rate, and eval regression checks after changes.
• Iteration loop: add new golden questions whenever new portfolio content ships.

• Over-retrieval: too many chunks increases cost and can dilute grounding — tuned k and chunk size.
• Missing evidence: system must refuse instead of guessing — enforced citation requirement.
• Stale index: portfolio updates require re-embedding — added a lightweight re-index workflow.