How the AI agent on this site works (and why it costs ₹0)

Recruiters rarely have time to read a whole portfolio, so I put an agent on this site that answers questions about my work. It's called Dumbledore; the chat window is only its interface. Here's how it's built.

The graph

The agent is a small LangGraph graph written in Python:

1. route: simple messages (hi, "how do I contact him?", "take me to his projects") get an instant reply without calling an LLM. That saves the free quota for real questions.
2. retrieve: hybrid search over this site's own content.
   • FAISS finds chunks with similar meaning, using one open-source embedding model, BAAI bge-small-en-v1.5: it runs inside the server itself (an 8-bit ONNX build, about 3 ms a question), with the same model on Cloudflare Workers AI as the backup, so every route gives the same vectors.
   • BM25 finds exact keywords.
   • Reciprocal Rank Fusion merges the two rankings.
3. scope: checks the question is about my work before any LLM sees it (more below).
4. rerank: keeps the best chunks. I A/B-tested a cross-encoder (bge-reranker-base) on the eval set: it tied RRF on hit@4 but ranked worse (MRR@4 0.885 vs 0.962), so the site keeps the RRF order.
5. respond: an LLM answers from those chunks only.
6. check: every sentence is checked before it reaches you.

Saying no: the refusal pipeline

An agent on a portfolio shouldn't write code, essays or opinions for anyone. Five gates keep it on topic:

• Rules: harmful requests, jailbreaks, personal questions and off-topic tasks are refused instantly.
• Meaning: the question's embedding (already computed for retrieval, so it's free) is compared with on-topic and off-topic example questions. Clearly off-topic means refused, still with no LLM call. A follow-up like "give me the link of the paper" is judged together with the previous question.
• Prompt: the model answers only from my profile, or replies with one exact refusal sentence. If a small model refuses a real question about me, a different model gets a second opinion.
• Output: the server holds each sentence until it's checked. Code or leaked instructions replace the whole answer; unknown links, phone numbers and other people's emails are removed.
• Cache: only clean, on-topic answers are ever cached.

On 76 attack and off-topic prompts there were 0 leaks. On prompts written after the rules (a held-out set), it refused 32 of 33, 24 of them before any LLM, with 0 false refusals on 47 real recruiter questions.

An agent that acts

Ask "open the LCSNet paper" and Dumbledore takes you there. After an answer it offers "take me there" buttons for the pages it was about. Where to go is decided by rules over an allow-list of this site's pages, so it can't be talked into sending anyone elsewhere.

Free models, with fallbacks

Free LLM tiers have rate limits, so one provider is never enough. Each question gets a difficulty level (fact, standard, reasoning), and each level has its own pool of free models across Groq, Gemini, Cloudflare Workers AI and OpenRouter. Before calling a model, the backend reserves its budget atomically in Redis, so many visitors at once never pile onto one model, and it trims older turns if a long conversation doesn't fit. The switch is invisible. Common questions are answered instantly from a bank of checked answers. Only when every free model is used up does the chat pause, until the limits reset.

Security in front, AI behind

• The browser never talks to the Python agent directly.
• A Next.js gateway checks for bots, validates input, rate-limits each visitor and enforces a daily budget.
• Only then does it call the agent, using a shared secret.

Memory that disappears

Chat history lives only in your open browser tab. When you close the chat or leave the site, it's gone. All I keep is a short AI-written summary, plus your details only if you chose to share them.

Measuring it

• Eval suites check that retrieval finds the right page, that answers contain the right facts, that guardrails and refusals hold, and that navigation lands on the right page.
• RAGAS scores faithfulness, answer relevancy and context precision, using a free model as the judge.