I build RL environments for frontier models and agents that take over real company processes.
RL environments and graders. Tasks mined from frontier-model failure modes, hidden scorers, replayable judge traces, reward-hacking checks. Plus the data infra behind them: distributed scraping workers, cross-region code corpora, dedup and quality filters.
Agents for business processes. A company brain that pulls finance, sales and ops data from 40+ connectors into one versioned store with access control. Consultant agents where every write waits for a human approval and numbers come only from verified tools. Every incoming stream (chats, calls, docs) digitized, every human decision point logged, and the recurring ones handed to agents. MCP servers, multi-agent swarms, agent memory.
Low-level performance. I profile C interpreters, parsers and compilers and make their hot paths faster. Every change is measured in instructions with callgrind, so the numbers are deterministic. Every change also passes the project's own test suite and a sanitizer build. Recent work: jq up to 2.8x faster, QuickJS array methods 1.44x, Duktape global variable lookups 2x.
ML in production. Recommenders shipped to 40M+ users, LLM moderation and retrieval pipelines, robot recovery loops, checkpoint-to-service delivery cut from 6h to 5m.
HL-EAI: A Multimodal Framework Enabling Emotional Reciprocity in Human–AI Strategic Decision-Making. Mozikov, Orekhov, Nasonov, Baltsat, et al., ACM Multimedia 2025. LLM agents and humans in dictator-game, GTBench and trolley-dilemma settings: how emotional cues change agent decisions, trust and cooperation.
Small, sharp fixes in other people's codebases, mostly parser, formatter and SDK correctness bugs.
| Project | Contribution |
|---|---|
| highlight.js ★24k | #4455 qualified generic type arguments in Java · #4453 where in Haskell GADT declarations · #4454 six-digit unicode ranges in CSS |
| pest ★4.8k | #1184 restore stack state after a failed sequence |
| agntcy/dir | #1892 register cloud KMS providers · #1891 avoid blocking on non-interactive key passwords · #1890 synchronize in-memory routing datastore |
| OpenFeature flagd (CNCF) | #2006 synchronize readiness state |
| OpenFeature go-sdk-contrib (CNCF) | #927 safely compare complex criteria values |
| libfyaml | #505 accept # in YAML tag URIs |
Open and under review: undici #5601 (reject SharedArrayBuffer-backed body views), oxc #25022, pygments #3234.
| Project | Change | Measured gain |
|---|---|---|
| jq ★36k | #3718 JSON parser fast paths | 2.25x on parsing |
| #3719 buffered output printer | 2.81x on -a, 1.5–1.7x on escapes |
|
#3716 @html escaping |
1.87x | |
#3717 explode |
1.65x | |
| QuickJS | #600 interpreter fast paths | array methods 1.44x, object literals 1.19x |
| Duktape ★6k | #2591 variable lookup, ToInt32, number-to-string | global lookups 2x, bit ops 2x, number printing 1.66x |
| lodepng | #236 PNG decode | 1.15–1.29x at -O2 |
Compiler and parser fixes: TinyCC 2d3eea20 chained designator initializers (pushed to mob) · Claude's C Compiler #332 11 bugs with minimal repros.
BSc in Applied Computer Science, ITMO – Best Graduate 2024 · Python, C++, PyTorch, JAX, MCP, Terraform