FreshVisit

The core lesson was that a browsing agent works better as a series of narrow, typed questions than as one open-ended prompt. On every screen FreshVisit asks Jev, TypeSafe's System One model, whether the goal is met, whether it can still be reached, and which of up to 200 elements to click next, and gets back probabilities and a confidence instead of free text to parse. I drove the browser through agent-browser's accessibility-tree snapshots rather than CSS selectors, which gave the model compact, meaningful choices like "button 'Pricing'" and made clicks one command. Running each journey in its own Trigger.dev container taught me retries, timeouts, and realtime progress feeds for long jobs that need a real browser. I also learned a lot about playing fair with other people's sites: staying on the start domain, stopping cleanly at CAPTCHAs, and never clicking anything that signs out, pays, or sends.

FreshVisit is a complete product I designed, built, and shipped on my own, from the landing page and pricing to team workspaces, PDF reports, and transactional email. It shows I can turn an AI capability into something a product manager, designer, or QA lead can use without knowing a model is involved. The work spans browser automation, AI decision-making, accessibility auditing, background jobs, media pipelines, billing, and auth, all in one codebase with unit and end-to-end tests. Persona docs and screen-by-screen walkthroughs shaped what got built, so the product decisions are as deliberate as the engineering.