Guided Portfolio Conversation
A portfolio guide with bounded context, public-source retrieval, and a real fresh-start button.
A guide that remembers the thread, stays grounded in public work, and lets visitors start fresh.
Jennifer Naomi Nguyen · conversational interface · Live · September 2026 · built with Claude Code and Codex
README · ARCHITECTURE · TECHNICAL · source
What I built
I wanted the portfolio assistant to feel less like a search box and more like a guide walking through an exhibit with you.
After visitors unlock the demo, they can ask about my research background, take a technical dive into a flagship project, and continue with natural follow-up questions. The guide remembers the active thread, shows whether it is carrying context, and can export either a clean transcript or a short HTML summary.
Remember the thread without turning it into evidence
I wanted follow-up questions to make sense without letting a visitor's earlier wording become evidence about me or my work. The complete active conversation stays in browser memory; each answer uses only a bounded summary and a small set of recent completed turns.
That history helps interpret references such as “How did that influence Attractor?” It is not evidence about me or my work. Factual answers are grounded separately in curated public portfolio pages and public-safe reference notes.
A conversation that can end cleanly
A visitor can type clear and start over without unlocking again. The command removes the browser-held conversation and any answer still in progress, while keeping the current page and one-hour session.
Apart from one explicitly contributed exchange, the application does not persist raw questions, answers, conversation summaries, recent turns, transcripts, or summary HTML in D1 or R2. A visitor who types contribute <1-5> after an eligible answer submits that single question and answer as an unreviewed evaluation candidate; nothing is stored verbatim without that command. It records limited operational details such as topic categories, prompt length, response mode, cited sources, and export kind and size. The Privacy & Safety Notice explains the visitor-facing data path.
Safe exports
A transcript is assembled locally from the conversation itself, so boot messages and navigation stay out of it.
An HTML summary is created only when requested. Generated text is inserted into a fixed, escaped template, and the downloaded page contains no scripts or tracking. The technical walkthrough covers the privacy and export boundaries in more detail.