CTO · PhD researcher in metaheuristic optimization
I work on combinatorial optimization, mostly vehicle routing, and mostly the parts where exact methods run out of road. My research puts machine learning and reinforcement learning inside metaheuristic search, so the algorithm learns which operator to apply and when.
I have also been writing software for 20+ years and leading engineering teams since 2019. Most of that work is private, so this profile is a poor sample of it.
Research. A CVRP research framework in Python: GVNS variants, learned operator selection, feature work on neighborhood prediction, and the instrumentation that makes search traces worth analyzing. Private while the papers and the dissertation are in flight.
Engineering. CTO at SoftHub. I run a multi-tenant payment gateway, a NestJS monorepo covering crypto, e-wallet and bank/card providers, event-driven over RabbitMQ. Next to it sits the sports data infrastructure: TCP socket feeds, SDQL/XML parsing, NDJSON pipelines, all of it on Kubernetes with Terraform. Private, and it is most of what I ship.
- CUDA by Example. Source code for the book's examples. 496 stars, the most-starred repo here.
- qwen-local. Hardware-aware local Qwen coding stack for Windows.
- claude-updater. Updates the whole Claude Code toolchain in one pass.
- Pointer-Networks-with-Pytorch. Minimal PyTorch implementation, left over from the TSP and RL detour.
- SublimeEncodingUtils and greek-srt-converter. Encoding tools. Greek text and UTF-8 have a long, unhappy history.
Python and C/CUDA for research, TypeScript and PHP in production. LaTeX for the parts that get reviewed.
Open to collaborations, especially the kind where the problem is hard and the people are sharp.