About
AI researcher.
Engineer. Builder.
I'm Andrey Chmerev, an independent AI researcher and engineer. I research, build and test new approaches in AI, and I treat shipped software as the experiment's result.
My day-to-day is applied research: new models, agent frameworks and local-first tooling, evaluated on real workloads and released when the results hold up.
What have I built?
Kekoso is a macOS app for on-device speech recognition. It dictates into any app with a warm model so there is no cold start, transcribes audio, video and YouTube links, records calls with two-track speaker separation, and exposes a local MCP server so Claude and other agents can transcribe and search recordings. It runs Whisper, Parakeet and SenseVoice models on Apple silicon, supports 99 languages, and never uploads audio. It is sold as a one-time purchase, which was itself an experiment in whether privacy-first AI can be a product.
THE FRAME is the process I extracted from building Kekoso and other projects alone with Claude Code. It is an MIT-licensed npm package that installs 33 commands, 10 agents and 5 safety hooks into a repository and gives a solo developer a six-phase loop: research, plan, build, review, ship, reflect. It keeps project memory between sessions, runs tests and quality gates automatically, and can drive the whole pipeline unattended.
What am I researching now?
- Agent workflows that stay predictable over weeks-long projects: how much structure an autonomous pipeline needs before it stops asking questions and still ships correct code.
- Local models and privacy-first AI on consumer hardware: what a Mac can run today at usable latency, and where the cloud still wins.
- The Model Context Protocol ecosystem: which servers actually earn a place in a daily workflow. Findings go into the MCP collections.
- AI-assisted development as a research method: treating every shipped product as an experiment with a written-up result.
How do I work?
Every idea starts as a research question with acceptance criteria. Then the pipeline plans the work in waves, builds it test-first, reviews the diff from six angles and ships a commit. I read the review, make the product decisions, and write down what I learned. The results show up here as products, open-source code and posts on the blog.