I build on-premise AI systems — for teams whose data has to stay inside.
Right now: Open Cradle, an on-premise agent platform, and the AI direction at Kyrgyz Single Window — systems for foreign-trade participants and the agencies issuing their permits.
Evidence
shipped & runningOpen Cradle
Regulated firms wanted AI on client documents, but professional secrecy rules out any cloud API. Cradle runs agents, retrieval, and risk triage entirely on the client's own hardware.
ResultEvery answer cites its source; an operator approves before anything leaves.
Government · 2025–nowSingle Window — permit workflows
An officer at one of the agencies that issues foreign-trade permits described a task eating hours of his team's week. Listening to that one problem produced ai.trade.kg.
ResultAI systems for foreign-trade participants and the agencies that issue their permits, built inside state infrastructure.
Product · 2025–nowCradle SmartNotes
A voice-first thought capture layer: local speech-to-text, wake-word screenshots, on-device grouping. The same on-premise constraint, applied to one person's working memory.
ResultNothing recorded ever touches a server that isn't yours.
How these systems are built
Rules decide, models explain.
Scoring, routing and thresholds are ordinary code you can read. The model writes the reasoning, not the verdict.
Every answer points at its source.
If a claim can't name the document it came from, it doesn't go out.
A person approves before anything leaves.
Approvals, edits and rejections are recorded — the log is what the system does anyway, not a report assembled afterwards.
How I work
clarity firstMost AI projects fail before the model is chosen — they start from the technology instead of the problem. So I start the other way: thirty minutes on what your organization actually struggles with, no slides, no jargon.
If there's a problem AI genuinely solves, you get a working prototype in weeks — built on your infrastructure, with your data staying where the law says it must. If there isn't one, I'll say so.