I build systems that make AI agents produce reliable work.
I'm Rom Iluz. 68 open-source repositories, 153 stars on CC10x, and the founder of Israel's largest agentic-engineering community. I build harnesses, workflows, and evidence-driven loops that turn AI coding agents from unpredictable one-shots into systems you can trust.
The work.
CC10x (153★), Hybrid-Search-RAG (64★), WhatsApp AI (46★), Memongo (29★), ClawMongo (29★), Auto-Pi (8★). Production systems with real users. Not tutorials — tools I use every day.
68 repos. Real users. Open source.
I don't write tutorials. I build systems I use every day, and other people start using them too. Here are the ones that proved the ideas work — not in theory, but in production, with real users, real stars, and real bugs that became real fixes.
CC10x
One router, nine specialist agents, sixteen skills, four evidence-gated workflows. The agent can't grade its own homework — an independent reviewer with zero context checks every plan and every build. Fail-closed gates, test honesty, bounded remediation loops.
View on GitHub →
Hybrid-Search-RAG
Production hybrid search that combines keyword and semantic retrieval. I built it because pure vector search misses exact matches, and pure keyword search misses meaning. It does both, and falls back gracefully when one path fails.
View on GitHub →
WhatsApp AI
AI agents that handle real conversations on WhatsApp. Not a demo — actual users, actual messages, actual edge cases at scale. The kind of project where you learn more from one week of production traffic than six months of benchmarks.
View on GitHub →
Memongo
MongoDB-native long-term memory for production AI agents. Vector search, episodic memory, and durable state that survives across sessions. I built it because file-based memory rots — agents forget, contradict themselves, and repeat mistakes they already made.
View on GitHub →
ClawMongo
A personal AI assistant that answers on WhatsApp, Telegram, Slack, Discord, and 15+ other channels — with MongoDB as the only canonical memory backend. Hybrid retrieval, automatic embeddings, and real agent memory tools. Fork of OpenClaw, rebuilt for production.
View on GitHub →
Auto-Pi
The same harness idea as CC10x, built on a different coding agent (Pi). Same system, different host — proving the pattern transfers. 10 packages, 59 skills, 10-step autonomous workflow. Zero bloat.
View on GitHub →How I think about agentic engineering.
The model is just capability. A better model doesn't fix unreliable work — it just produces unreliable work faster. What makes an agent trustworthy is the system around it.
The harness controls what the agent sees and does. Context, tools, authority, skills, memory, evidence requirements. Curated, not dumped. Enforced structurally, not by asking nicely.
The workflow turns judgment into an executable path. Router → workflow → agents → skills. Your team's engineering decisions become a contract the agent must follow — not a suggestion it can ignore.
The loop uses evidence to decide what happens next. Not retries. Convergence loops. Claim → evidence → verdict → next state. The system advances, repairs, re-plans, stops, or escalates based on proof — not the agent's confidence.
I built CC10x and Auto-Pi to prove this works on different hosts. I built Memongo and ClawMongo because agents without durable memory repeat their own mistakes. I built Hybrid-Search-RAG because retrieval quality is the difference between an agent that helps and one that hallucinates.
Want to work together?
I'm available for keynotes, workshops, and advisory engagements. I also build things in the open — follow the work, fork the repos, or reach out directly.