AI Digital Twin

The chat button in the corner of every page opens my AI digital twin. It answers recruiters' questions about my experience, skills and projects in the first person, and lets them leave their email so I can get back to them.

Where it started

I learned the pattern in Ed Donner's Complete Agent and MCP course. My first version was a Python app with a Gradio chat window and an OpenAI model. It read my LinkedIn profile and a short summary, and it had two tools that send me a Pushover notification: one when someone shares their contact details, and one when it gets a question it can't answer.

Fitting it into this site

Instead of hosting a separate Python app, I ported the agent to TypeScript and built it into this portfolio, so it shares the same repo, tests, deploys and domain.

• A Next.js route handler on Vercel runs the agent with the Vercel AI SDK, which streams the answer and handles the tool calls. The model is Claude Haiku.
• The prompt is built from my LinkedIn profile, a summary and the project pages on this site, so it stays on career topics and records any question it can't answer instead of making something up. A small script turns the project pages into the twin's knowledge, and a test fails if I forget to run it.
• The chat window is a React component that loads only when you open it, so it adds nothing to the page load. On phones it opens full screen.

Keeping a public endpoint safe

Anyone can call an AI endpoint on a public site, so most of the work went into protecting it, cheapest checks first:

• Two kill switches: an environment variable and a flag in Upstash Redis.
• A same-origin check, Vercel BotID and strict limits on message size.
• Per-visitor and site-wide rate limits in Upstash Redis, plus a Vercel Firewall rule.
• A daily cap on notifications and a hard spending limit on the API key.

How it's checked

• Vitest covers the prompt, the tools, the rate limits and the API route.
• Playwright tests the chat window against a mocked stream, including keyboard use and an axe accessibility scan in both themes.

What I learned

Calling the model was the easy part. Most of the effort went into the prompt, abuse protection and cost control, which is what turns a demo into something I'm happy to leave running on a public site.