05Multi-agent · Local-first
The memory is markdown in Git, so it outlives the runtime that reads it.
My own assistant for memory, notes, personal finances and calendar, which I talk to over Telegram and a web interface. Assistants that advertise memory keep it somewhere you cannot read or take with you. I wanted the opposite: delete the app tomorrow and everything it knows still opens in a text editor.
So its memory is a folder of markdown notes kept in Git, and the application is a replaceable layer on top. Anything it cannot undo, like changing a financial record or sending a message, waits for my approval first. Backups run on their own, and it all costs under $20 a month.
The tools, the architecture and the engineering numbers, one line each.
01MLOps · Time series
Over 100,000 SKU+Store combinations, forecast weekly by a routing policy that picks a model per cluster.
View full case study: ML demand forecasting in production02Data platform · Generative AI
Weekly public data, parsed into a 52-week forecast the buying team takes into supplier negotiations.
View full case study: Market price intelligence platform03Applied AI · Guardrails
A Telegram assistant for questions about sales, inventory and purchasing. Every figure it gives comes from a query it actually ran.
View full case study: AI assistant for business data04Full stack · Internal product
Analysts used to work out every inter-store transfer by hand. The engine proposes what moves where, and they run it themselves.
View full case study: Stock rebalancing engine