I build backend systems in Node.js and TypeScript, and I run a business unit at Onix, a Ukrainian software company that works with clients in the US and Europe. I've shipped more than 30 production apps and still spend a good part of my week in code.
Most of my recent work is putting LLMs into products where a wrong answer is expensive. That usually means narrow model calls with deterministic checks around them. One example is a text-to-SQL layer over a multi-tenant analytics database. It makes three model calls (route, generate, present). A validator checks the SQL as a syntax tree before it runs, queries are read-only and scoped to one tenant, and the model never writes a number itself.
Other things I keep coming back to:
- RAG pipelines with adaptive query routing
- Agent tooling with MCP, Mastra, LangChain, and the Claude API
- Enterprise integrations and multi-tenant SaaS
- Domain-driven design and hexagonal architecture
Storewalk goes through an online store the way an AI shopping agent would and shows where the agent gets stuck before it can buy. It also checks which of the new agent standards the store supports. We built it at Onix. It's free, and we've just started testing it.
Quixli is one editor for the documents you send to clients, like proposals, statements of work, technical docs, and reports. Each document lives at a link, and you can see when it was opened.
I write on LinkedIn about engineering around LLMs. The latest piece is "The Text-to-SQL Nobody Demos", and I'm working on posts about harness engineering. I also write on Medium.