Yehor Kaliberda — AI Systems Architect · Aarhus, DK
I build autonomous AI systems that do real engineering work and prove it: merged pull requests, machine-checked verification, audit trails you can replay.
Shipped — CallMed AI
CallMed AI is my one-person agency for autonomous Python codebase audits. The proof is public: real pull requests, reviewed and merged by maintainers who had no reason to be kind.
Deterministic AI-code verification · shipped
Checks every import, symbol, and API call an AI assistant writes against what's actually installed — no model in the loop, so the same input always produces the same output. Open-core CLI published to PyPI and npm, with signed build provenance behind every release.
Autonomous security remediation · shipped
Runs five scanners in parallel, triages findings, generates one focused fix per finding, and verifies each through a three-gate deterministic pipeline before opening a draft PR. Local-first: the code never leaves the client's machine except the fix prompt itself.
Autonomous type annotation · shipped
Parses a Python repository into a typed AST dependency graph, then writes PEP 484 annotations under collision-free file locks. Strictly additive: logic, comments, and docstrings stay untouched. Delivered as a single reviewable PR with 0 pyright errors.
11
Provenar PRs merged across 4 open-source projects
217
FixProve tests in CI, zero LLM tokens in the check itself
565
Patchward tests passing, at 91.20% coverage
0
pyright errors on every Provenar delivery
In build — 2026
A multi-agent system that rediscovers physical laws from raw data, then proves them in Lean 4. Currently hand-formalizing differential geometry from scratch, one tensor at a time — nothing counts until the proof compiles.
Read the build notes →A productivity agent that watches your screen, redacts anything private before inference, and runs entirely on a 4 GB consumer GPU.
Read the build notes →A personal editorial project, governed the same way the code is: every claim needs a source, nothing publishes without two separate human signatures. Early stage — nothing public yet.
A recursive, self-correcting document built with several AI systems — every claim sourced, every wrong correction logged as a wrong correction rather than quietly fixed. Under active revision, running a small human pilot before it goes further.
A live, sealed forecasting ledger. Two AI models, Astra and Fable, lock probabilities before deadlines and are scored on what actually happens — outcomes observed by machine where possible, every entry hashed at lock time. Paper scores, nothing staked.
Enter the observatory →Philosophy
Contact
I take a small number of engagements at a time. Verify your identity through the portal, or send the scope directly and I reply within a day.