Indie AI/ML engineer in Toronto. I build AI things, with AI things, ship them before they're perfect, and write about what I learn along the way.
-
Making your AI Evals cheap and precise.
I built U/=22A8, the platform for models that are tailored specifically for AI Evals, replacing LLM-a-as-judge with DLM-as-a-judge. Evals for your AI stack, without generating output tokens. 23Γ cheaper than the typical LLM-as-a-judge β more accurate, lower latency, deterministic. Benchmarked against Claude Opus 4.7, the only judge that matches our quality β 195Γ cheaper, 19Γ faster. benchmarks β. If cheap evals is something you're interested in or want to rant about LLMs for Evals, you can chat to me right this moment - t.me/yanchynskyy; I'll tell Claude to take a break while we chat about it. -
Measuring your code's βοΈvelocityβοΈ.
I build a CLI tool in Rust dubbed Terminal Velocity that turns your git history into an insight factory about how effective you're shipping. Runtv --repo <your-repo>and let me know what you think. It measures: code flow, commit size, thrashing, excision, cadence. The tool will tell you the real true story, not imagined arbitrary delivery metrics. The real measure of delivery is flow, the terminal state of your engineering velocity. -
Adopting AI-augmented ADHD-friendly productivity system.
I maintain a plugin for Claude dubbed Stream of Consciousness. It's a Claude Code plugin for brains that don't do well with to-do lists. Things flow in, decay over time, and either get resolved or restreamed. No projects, no tags, no priorities, no need to get organized, just vibes and a clock. -
Learn about Explainable AI.
I run interference substack with a friend of mine. We write about xAI. For example, check out A Practical Guide to Explainable AI. -
Shipping AI things or with the help of AI.
I think out loud about AI, building, and shipping at high velocity with the help of AI and small team. tarasyanchynskyy.substack.com is the medium.
Building in public, one bit at a time.