Ask my AI: Portfolio Assistant

The problem

Recruiters and hiring managers skim portfolios quickly and often have one specific question. A chat assistant can answer it directly, but only if it never invents experience, stays on topic, resists prompt injection, and shows where each answer came from.

How it works

What I did

• Grounded every answer in the same content file that renders the site, so the assistant and the pages never disagree.
• Streamed answers through the OpenAI Responses API, fetching live GitHub or Medium data only for questions that need it, and restricting web search to approved domains.
• Had the model end each answer with the IDs of the sources it used, then stripped that line from the stream and kept only IDs on a server-side allowlist, so it can never introduce a link of its own.
• Linked each source to the exact role or case-study section, which scrolls into view and is briefly highlighted.
• Protected it with input moderation, request size limits, per-visitor and daily rate limits, and disabled response storage.
• Built an evaluation suite that replays real questions, including multi-turn follow-ups, prompt-injection attempts, and questions with no published answer, and checks the answer text, the cited sources, and that every link lands on a real anchor.

Results

• Passes all 34 evaluation cases, including follow-ups where the sources must change with the question.
• Says when something is not published instead of guessing, and declines off-topic and injection attempts.
• Every source link opens the exact part of the site the answer relied on.