AI Portfolio Assistant

The assistant lives at /assistant and lets anyone have a real conversation about my background. It knows about my CV, my skills, my personal story, and every single project in my portfolio. Ask it whether I'd be a good fit for a backend Python role, what games I've built, or whether I have CI/CD experience — it retrieves the relevant information from my actual documents and gives you a grounded, cited answer.

The architecture is a full RAG pipeline built as a Flask blueprint. At startup, it indexes two types of knowledge: static Markdown files (my CV, skills breakdown, and personal narrative) and live project records queried straight from the PostgreSQL portfolio database. That second part was a deliberate design choice — it means the assistant always reflects the current state of my portfolio. Every time I add or update a project through the admin dashboard, it's automatically available to the assistant on the next deploy, with no manual knowledge base maintenance required.

Embeddings run entirely locally using sentence-transformers/all-MiniLM-L6-v2 — no API key, no cost, no latency from an external call. ChromaDB stores the vector index on disk so it only rebuilds when needed. At query time, the four most semantically relevant chunks are retrieved and injected into the prompt alongside the conversation history before the request goes to Llama 3.3 70B on Groq.

One of the details I'm most pleased with is the smart source citations. When the assistant answers, it doesn't just say "source: projects" — it links somewhere specific. A question about a particular project links straight to that project's detail page. A question about Flask links to /projects?tag=flask. A question about data work links to /projects/data. A broad question falls back to /projects. This required building a tag frequency system across retrieved chunks and a GENERIC_TAGS exclusion list to stop tags like "python" or "programming" from winning over meaningful ones like "flask" or "machine-learning".

The chat UI is built with Bootstrap 5 — no page reloads, everything via fetch(). It renders Markdown, persists the full conversation in localStorage with a clear and restore flow, and ships with twelve starter prompts covering the range of questions a recruiter might actually ask. Rate limiting is enforced per IP at 15 requests per minute and 1,500 per day using flask-limiter, with context-aware error messages depending on which limit was hit.