Staff Software Engineer at Applied Materials and co-founder and CTO of CarrotOnTheHorse. Fifteen years of shipping software, the last few spent turning generative AI into tooling people actually depend on.
Messy real-world data, made useful. The same instinct at enterprise scale and in a barn.
Co-founder and CTO
Founded 2026, one customer live
Ask a trainer or barn owner where a horse’s records live and you hear the same answers: a binder in the feed room, a truck full of texts, one person’s head. The information exists. It is just trapped.
CarrotOnTheHorse takes what a barn already produces, such as vet PDFs, farrier sheets, photos and voice notes, and turns it into a structured memory of every horse: source-backed records, tasks, and answers anyone on the barn can ask for. Import, not input.
I lead AI, architecture and engineering: multimodal document parsing, entity resolution, retrieval and agent workflows, and human-in-the-loop review for animal-health data where a wrong answer matters. Most of it runs on Cloudflare.
TypeScriptCloudflareLaravelReactGemini, OpenAI and Anthropic APIs
Enterprise GenAI for a very large engineering organisation: chat apps, retrieval pipelines and agent workflows on OpenAI, Anthropic and Gemini models, running across GCP and Azure.
GoPythonTypeScriptJavaElasticArangoDBPostgresKubernetes
A GenAI and wearables startup predicting musculoskeletal pain before it starts. I built most of the platform, hired and led the engineering team, and wrote the scoring that turns Apple Watch activity into wellness scores, recovery plans and a chat assistant that triages discomfort.
An autonomous coding agent that picks up GitHub issues and Discord requests, works in isolated checkouts, and reports back without a human driving it. Most of the effort goes into reliability: authorised actions only, no duplicate runs, and recovery when a job dies halfway. Runs on Cloudflare end to end, databases included.
TypeScriptCloudflare WorkersCloudflare storage and databases
A native woodland game in C++ where dogs chase squirrels through terrain I generate myself. An excuse to write rendering, physics and navigation from close to the metal, and to build a world of my own.
C++20SDL3MetalJoltRecast
Can keyword and vector retrieval run well together on a laptop or a small home server, without a cloud bill? A Rust playground for finding out.
Rust
Field notes from running AI in production, at work and in the barn.
Gemini 3.5 Flash-Lite shipped yesterday. We had it in our pipeline by lunch, red tests by afternoon, and a coding agent proposing a time-travel database by dinner. A day-one field report from Carrot on the Horse.
Agentic coding tools are the most empowering thing to hit engineering in a decade. They’re also a quiet wrecking ball. Both are true at once, and pretending otherwise is how teams get burned.
I have been writing software since school, for pay since 2012, and I still ship from the first line of code to the production rollout. The route ran through e-commerce shops in Moscow, internal tooling in St Petersburg, and now Santa Clara, where I have spent nearly a decade at Applied Materials.
My sweet spot is messy, real-world data: documents, code, images, voice, conversations. I like making AI structure it into something useful, secure and production-grade, and I care most about the boring-but-hard parts that make it trustworthy rather than merely impressive in a demo.
The question I spend most of my time on now is whether any of it paid off. Getting AI into production turned out to be the easy half. Proving it moves the needle, with adoption and hours saved rather than anecdotes, is the harder one.
Away from software, there’s my dark Gypsy Vanner, a backyard goldfish pond, and a Mazda MX-5. I also play Dota 2 and World of Warcraft. Between the horse and the pond there is always something to tinker with, and a fair amount of scattered paperwork to structure.
Old friend, former colleague, fellow builder, or just curious? I’m always happy to talk shop on GenAI, agents, measuring whether AI paid off, or horses.