AI Safety Engineer • Quantitative Developer • Open Source Builder
I build systems that think, learn, and protect.
The world's first Safety-First Guardian Layer for LLM code-fixing agents.
OpenCode writes fixes. Coding Tentacle analyzes, blocks dangerous changes, requires human approval, and learns from every run. Safety VETO + SkepticBrain + Bayesian Trust + 18 bug types. 68 Python modules. 100% open source.
→ github.com/nessos666/coding-tentacle
Three production-ready tools — no API keys required, MIT licensed, CI-tested on Python 3.10/3.11/3.12.
|
148 key-free academic sources searched in parallel. Give it a topic, get a PRISMA dossier with ranked, verified papers.
→ Repo |
40+ sources in parallel — web and academic. Keeps working when every other search tool fails.
→ Repo |
Turn AI conversations into durable memory. Quality gates decide what is worth keeping — and what is noise.
→ Repo |
I don't just write code. I build systems with intelligence.
| 🔬 Research First | Deep literature review before architecture. 12-layer research sweeps. I read the papers. Then I build. |
| 🛡️ Safety-First | Every system I build has a VETO layer. Nothing executes unless it's safe. No exceptions. |
| 🧠 Self-Learning | My systems learn from every run. Bayesian trust calibration. Experience consolidation. Memory bridges. |
| 🔧 Full Stack | From kernel-level monitoring to AI agent orchestration. Python, Rust, Bash, C++, Go, Ruby, JS/TS, Java. |
| 📊 Evidence-Driven | I benchmark everything. If it can't be measured, it can't be improved. 100-issue reality tests as standard. |
| 📦 Open Source | Everything I build is public. MIT license. No paywalls. No black boxes. |
| Project | Description | Stars |
|---|---|---|
| coding-tentacle | Safety-first guardian layer for LLM code fixing. VETO + SkepticBrain + Bayesian Trust. | ★0 |
| context-bench | Self-learning Claude Code hook that auto-injects project context into every session. | ★0 |
| multi-rep-set | Multi-language code checker. Bash, Python, JS, JSON, YAML, Markdown + System. Zero dependencies. | ★0 |
| Project | Description |
|---|---|
| topstepx-api-health-monitor | 9 automated checks → Trust Score 0-100. Contract rollover, token expiry, bar quality, latency trending. |
| rithmic-api-health-monitor | File-based health monitoring for Rithmic API — 10 checks, no own connection. |
| api-health-trust-system | Generic weighted Trust Score for any REST API. Pluggable checks, Telegram alerts, atomic JSON export. |
| Project | Description |
|---|---|
| nq-strategy-builder | IS/OOS/Holdout backtesting with NLP-driven strategy composition. 158 files, 60 tests. |
| quant-tools | Ornstein-Uhlenbeck MLE fitting, simulation, bias correction, volatility estimation. |
| baustein-tester | Isolate and rank signal concepts (FVG, OB, Sweep, Breaker) before strategy building. |
| Project | Description |
|---|---|
| chart-vision-mcp | Local TradingView chart analysis via OpenCV + Tesseract OCR. |
| tv-watch-agent | Automated TradingView chart watcher — CDP-based screenshot agent. |
| wissenschaft | 148 key-free academic sources in parallel → PRISMA dossier. Citation snowballing. |
| sucher-1000 | 40+ web & academic sources in parallel. Multi-engine fallback, FTS5 archive. |
| extraktor | Extract durable knowledge from AI chats into Qdrant. 5 quality gates. |
| graveyard | Archived experimental projects. |
AI Safety ██████████ Coding Tentacle, Context-Bench
Infra/Monitor ████████ API Health Trust System, TopstepX, Rithmic
Quant/Finance ███████ NQ Strategy Builder, Quant Tools, Baustein Tester
Tools/Utils ██████ Multi-Rep-Set, Chart-Vision, TV Watch Agent
🦑 Built with curiosity, shipped with care. Hungary → Germany. Bilingual DE/HU. Open to AI Safety & Infrastructure roles in Budapest.