Kelvyn Abreu
Custom AI applications, directed from first prompt to final release.
Magik Table
A self-hosted database platform: tables, kanban boards, forms and dashboards on your own server.
- 7 days, first prompt to version 3.2
- 55 prompts logged, word for word
- 3 web apps on one database
- 203 automated tests
The brief
A flexible table database that runs locally and keeps its files local, with kanban, charts and forms included rather than added on.
The build
It began as a single desktop app. On day two it was redesigned as separate web apps: a server console for IT, a client for teams, and later a kiosk for employees and the public, all on one database.
The release
Version 3.2 runs many independent organizations on one server, with a role hierarchy, SQLite, PostgreSQL or MySQL storage, backups and printed manuals.
Built with: Python, FastAPI, SQLite, PostgreSQL, MySQL, JavaScript, Claude Code
Streame
One live stream in, every platform out, run from a single control panel on your own server.
- 4 days, first prompt to 1.0 beta
- 60 prompts logged, word for word
- 88 automated tests
- 3 platforms connected by sign-in
The brief
Take one live video source and send it to many destinations from a self-hosted server, controlled through a web panel in the style of a hardware encoder.
The build
A Python web service with a media engine for incoming streams and one FFmpeg process per destination, each retried by itself if a connection drops.
The release
Companies with their own users and white-label branding, live previews and audio meters, an OBS dock, and sign-in to YouTube, Facebook and Twitch for titles and descriptions.
Built with: Python, FastAPI, FFmpeg, MediaMTX, OAuth, JavaScript, Claude Code
X-Ercise
Your training coach: plan, activity and recovery in one place.
- 6 AI providers, cloud or self-hosted
- 30 tools the AI coach can use
- 15 dashboard widgets
- 3 layers keeping the coach on topic
The brief
A self-hosted training coach that brings workouts, recovery and goals together and recommends today's session, for each person on their own private account.
The build
A Python web app with one rules-based engine for the daily recommendation and one validated layer for every change, so the AI can only save what a form would accept.
The release
A coach that each person connects to their own AI or a shared one, reading only their data and acting only when asked. It installs like a phone app, with dashboards people build themselves.
Built with: Python, Flask, Waitress, SQLite, JavaScript, Web Push, Claude Code