Pre-Visit

Pre-Visit is the demo site I built for an AI-driven patient-intake product, and it's the live link every cold email from the Opportunity Intelligence Engine actually points to. The site itself is a Next.js 15 app: an intake chat panel showing what a patient's pre-visit conversation looks like, a booking calendar, phone and identity verification steps, and a contact path, all built to be the thing a practice owner clicks through in under a minute and immediately understands.

The part I'm most glad I built properly is the demo video. The brief was simple: show a patient's whole flow, then cut to what the doctor actually sees. I built it as a frame-by-frame renderer rather than reaching for timeline-editing software: each scene is a self-contained HTML file exposing a pure renderFrame(progress) function (0 to 1), nothing in it depends on Date.now() or a real CSS transition, so the same progress value always produces the same pixels. A Puppeteer script loops through progress values, screenshots each frame, and ffmpeg encodes the sequence into a silent MP4 per scene before crossfading them together and muxing in an audio track trimmed to the assembled video's exact runtime via ffprobe.

That determinism is what made iteration survivable. Early passes had real, fixable problems: reveals timed to a flat 0.5 seconds regardless of how much text there was to read, hard cuts between scenes that read as a glitch, a closing line whose word-by-word reveal finished right as the scene ended with no time to actually read it. Every fix was a change to one scene's pacing windows, not a re-shoot, because re-rendering a single scene and re-running the assembly script always reproduces exactly the same result. The motion itself went through a real upgrade pass too: the logo draws its stroke on instead of fading, cards drop in with a slight 3D tilt and spring overshoot, the doctor's queue view gets a moving highlight sweep instead of a static pulse, and each side of the story opens with its own full-bleed chapter wipe so the two-sided flow reads as two explicit parts rather than one undifferentiated scroll.

I built this pipeline specifically because the equivalent one used for an earlier demo video was never committed anywhere, only the rendered output survived, and rebuilding from a lost process is not a mistake worth repeating. Everything, scenes, render scripts, and the final output's source, lives in the repo now.