I grow social accountssocial accounts from scratch, then build the software that keeps them growing.

A person talks for twenty minutes. That transcript becomes the week, written in a voice profile the system holds and checks itself against. Quality gates sit between the model and the feed: length caps, automatic re-rolls, and a publish hold on anything that reads wrong.

Built with Node, the Anthropic API, headless-browser rendering for the image pieces, and a scheduler that posts on a collision-free calendar.

Status Running daily as a client's production system. I built it; the tool is theirs.

Showing up in other people's comments is how small accounts get found. It also eats an hour a day. This drafts the comments in your voice and stacks them for review, so the hour becomes about two minutes.

The part that matters: it never presses Post. It fills the box and you send. It reads posts the way any signed-out visitor does, never through your login, and it does nothing to hide what it is. Platforms restrict accounts for automated activity, so the automation stops at the draft and a person stays on the button.

Built with Python, the Anthropic API, a browser extension, and official platform integrations.

Status Live and in daily use on a client's accounts. I built it; the tool is theirs.

Motion graphics are usually the bottleneck: one designer, a queue, and a week of turnaround. Here the shots are written as code, so the house style is a file rather than a habit. Type, colour, spacing and timing are rules the render obeys, and a shot that breaks one fails its check instead of shipping.

Then the whole setup packs up and installs on someone else's machine. That is the harder half. What makes an AI workflow good lives outside the project folder, so copying files moves nothing; the kit captures the rules, the memory and the environment, and gates the install on a frame-by-frame acceptance test.

Built with Remotion and React, rendered locally through ffmpeg on a pinned toolchain. No render fees and no per-seat cost: the framework is free for individuals and teams of up to three, and licensed above that.

Status Producing finished shots. The transfer kit is written and waiting on its first install.

A clinic with three sites and thirteen thousand paper charts cannot stop seeing patients to type them up. So the phone does the reading and the staff do the confirming. In the test that convinced me it was worth building, a handwritten prescription came back correct in one pass, the one ambiguous quantity was flagged for a human instead of guessed, and the field that was blank on paper stayed blank on screen.

Nothing saves on its own. A person confirms every record, and the photograph stays attached to it.

Built with a Vite app on Cloudflare, Postgres in the Singapore region, and the Anthropic API for the read. Access logging, encrypted backups and role limits are part of the build, not an add-on.

Status Demonstrated and scoped. Screens here run on invented patients.