Founder, Heretic OS — the layer an AI agent runs on.

Not a bigger brain. The brain stem underneath it: a memory that hands back the original word for word, a live map of the codebase the agent thinks with, security inside the loop, and process gates a stage cannot close without.

12 years of industrial engineering in regulated environments — oil & gas, FMCG, large-scale construction — where an unlogged decision is a finding and a wrong one is an incident. That is the half this field is short of. Chemnitz, Germany.

I ship systems with tests, evals and audit trails. Not demos.

Memory folds by address and returns the original byte for byte — 100% recovery across 1,008 live folds, zero loss and zero false alarms. Against the field's best-funded open memory system on LoCoMo: 0.685 against 0.438 under a strict independent judge, on 10.4× fewer write-time model calls (656 against 6,836, counted on the same corpus). Runs in the cloud, on a laptop, or with the network cable cut.

172,997 lines of product code · 549,394 lines of tests · 13,197 tests passing · 2 inventions filed with the USPTO

Closed-source commercial system — architecture tour in the repo.

An 11-pass Tree-sitter pipeline that turns a repository into something an agent can query: who calls whom, where a value travels, which test covers which line, and what breaks if you touch this. Full depth in ~15 languages, call graph across ~29, symbol inventory across 39.

95.8% call-graph precision against PyCG's 96.2% — on PyCG's own public 119-case suite, both engines run by us under the same scorer, rather than quoted from either paper. One false positive in 60 edges; the reference tool had two.

153,923 entities · 920,877 edges · one pass over a 2,148-file tree

Physics-based plant simulation (Weibull/IEEE 493, ISO 281, Arrhenius, TEMA) into explainable ML — IsolationForest + SHAP detection, XGBoost RUL with split-conformal intervals — into a risk-adjusted decision engine with ISA-18.2 alarm management. Approved decisions act back on the plant, and every repair is booked against its own run-to-failure counterfactual.

Authority-ranked human oversight per EU AI Act Art. 14: two of four options need a Production Manager or a Plant Manager, and if nobody answers inside the response window the recommendation applies under its own named identity — never under a human's.

1,300+ tests · 78% coverage · live demo — open at any hour, audit trail populated

Lemon detection with a leakage-safe pipeline, plus an opportunity layer the brief did not ask for: OOF-calibrated mining of undervalued low-risk vehicles with Wilson intervals.

F1 0.34 → 0.42 · 92–98% holdout on the opportunity layer

A public series, each article published with its method and an evidence table beside it — written to survive technical and legal review before it goes up.

• Telemetry is not evidence — a log a system writes about itself is testimony, not proof
• A tamper-evident, append-only audit trail for autonomous agents — a defensive publication
• Integrity proof and erasure proof in the same ledger — the GDPR pair, solved once
• Wired, suspect, unwired — code that exists, passes its tests, and is connected to nothing

Full series: heretics.dev/research

Python FastAPI scikit-learn XGBoost PyTorch SQLite TimescaleDB Redis Streams Docker MCP Tree-sitter React/TypeScript

🌐 heretics.dev · 👤 muratsuer.eu · 💼 LinkedIn · ✉️ murat@muratsuer.eu