Portfolio AI Assistant

Portfolio websites are static. Visitors who want specific information have to read through multiple pages or fill out a contact form and wait. This creates friction for recruiters and potential clients evaluating multiple candidates.

Built a custom RAG pipeline with content-type-aware chunking (paragraph-level for bio, per-project for projects, category-level for skills, QA pairs for FAQs). Implemented confidence scoring with fallback to contact form when retrieval relevance is below threshold. Used streaming responses for perceived performance.

92% of questions answered from knowledge base with zero hallucination incidents in 3 months

• Average response time: 3.2 seconds (target was 5)
• 92% of questions answered from knowledge base
• Zero hallucination incidents in 3 months of operation
• 65% of visitors engaged with the assistant
• 15% of engagements led to contact form submission
• Top questions: technologies, availability, pricing

Tradeoffs and lessons

RAG quality depends on chunking strategy — content-type-aware chunking dramatically improved retrieval precision. Confidence thresholds prevent hallucinations better than prompt instructions. Streaming responses transform perceived performance — a 4-second streaming response feels faster than a 2-second batch response.