Second-year software engineering student at Taras Shevchenko National University of Kyiv. I work on applied AI: agents that act on a computer under the user's control, search over private documents, and benchmarks that turn "it works" into a number.
Open to junior AI engineer roles and internships, remote. LinkedIn · stepanovgleb10@gmail.com
| Project | What it is | Measured |
|---|---|---|
| Halo Bot | Desktop app where AI assistants do real work on your PC, each in its own Windows sandbox, behind an approval gate and an audit log | On an 11-task benchmark a local 4B model went from 7 to 10 solved tasks (PR #14) |
| Glossa | Ask your documents a question and get the answer with the page it came from. 35 languages, runs offline | Retrieval pipeline in about 300 lines, no LangChain |
| vybir | Runs an AI decision model on ordinary PCs and hands the cases it isn't sure about to a person | Per-class calibration cut spam that got through from 23.8% to 3.5%; 183 tests |
| skillrot | Shows how much of an AI agent's context its skill library takes up | Weekly CI audit of 5 public libraries |
| VaultDrop | Offline Windows backups that read back and verify every file | Installer and CLI in v0.2.0 |
| RoadForge | Self-driving demo on real OpenStreetMap roads, in the browser | Live demo; 0 collisions in 20 benchmark runs |
Right now I'm training my own 158M-parameter language model from scratch on Ukrainian, English and code. That repo stays private until a checkpoint is worth sharing.
I write code together with AI coding agents (Claude Code, Codex) and hold it to the same bar as anything else: tests, CI, and a benchmark before I call a change an improvement. The numbers above come from those tests and benchmarks, not from estimates.