I build AI agent runtimes you can actually trust in production.
Federated AI for Frappe / ERPNext · governed agent harnesses · full-stack from kernel to UI
I work on the layer between large language models and real businesses: the harness — the governance, memory, and tooling that decides what an agent is allowed to see, remember, and do. Models are commodities; the harness is where trust lives. I also run the systems underneath it: distributed data engines (Redis, Valkey, PostgreSQL, MongoDB, Kafka, vector DBs) across cloud and bare-metal, provisioned with Terraform on RHEL/Ubuntu.
- 🛡️ Muster — an open-source agent harness you can audit. Token-waste ledger, leak-proof scoped memory, eval-gated learning, a never-wedge compactor, and one governed envelope for Telegram / Slack / Discord / WhatsApp / Teams / web. 20+ LLM providers, zero lock-in. TypeScript · MIT.
- 🤖 frappe-agent — a Frappe/ERPNext agent plugin for Codex & Claude Code: DocType design, customizations, bench, SQL, and ERPNext workflows. My most-starred work.
- 🗄️ OSS Manager — a control plane for self-hosted data engines: lifecycle, sizing, preflight, and ops automation across Redis · Valkey · PostgreSQL · MongoDB · Kafka · Qdrant · Milvus on AWS, GCP, and bare-metal — Linux (RHEL/Ubuntu), Terraform-provisioned, with a federated support copilot on top. Go · Python.
- 🪟 Clawglass — a native macOS manager for OpenClaw instances, with glass-inspired setup flows. Swift.
- 📚 AI Systems Engineering — course material: from first principles to production agents.
Self-improving agents are easy. Provably governed agents are the hard part — every memory scoped, every learned skill behind an eval gate, every token on a ledger.
I re-model the best ideas from OpenClaw, Hermes, and pi.dev — then fix the failure modes their own issue trackers document, and bundle it leaner. Optimize first; ship the audit trail with the feature.
Languages
Data & streaming engines
Cloud, infra & ops
AI & app platforms
Currently going deep on: governed agent runtimes, agentic RAG over ERP data, and shipping Muster toward its first release.




