A CV that interviews back: one dev's overengineered portfolio

What it does

This is a personal portfolio site with an AI avatar named “Santi” that visitors can chat with via text or voice. The avatar answers in first person about Santiago’s experience, backed by a retrieval system that pulls from his case studies and CV content. A private /ops dashboard surfaces real production traces, costs, and security events.

The interesting bit

The project is essentially a demonstration that the skills listed are the skills demonstrated. The chatbot runs a six-layer prompt-injection defense with real-time jailbreak email alerts, 71 automated evals that gate CI deploys, and a closed-loop where low-quality production traces auto-generate new test cases. The voice mode costs roughly $0.25 per session. Even the architecture diagram is GSAP-animated with narrated audio.

Key highlights

• Hybrid RAG pipeline: pgvector semantic search + BM25 full-text + Claude Haiku reranking
• 71 evals across 10 categories, ~70% deterministic, ~30% LLM-as-Judge
• Private LLMOps dashboard with 8 tabs: costs per component, security funnel, voice latency P50/P95
• Voice mode shares the same RAG pipeline via OpenAI Realtime API WebSocket
• Build chain includes prompt regression testing, RAG diagnostics, and SEO validation

Caveats

• The README is thorough but the repo itself appears to be a personal portfolio; reuse beyond reference is unclear
• Several features (jailbreak alerts, voice mode, ops dashboard) depend on external API keys and services that aren’t free

Verdict

Worth studying if you’re building production RAG chatbots and care about evals, observability, or prompt security. Skip it if you want a drop-in component — this is tightly coupled to one person’s content and infrastructure.