Skip to main content

Start Here

If you only spend 5 minutes here.

Curated front door. 8 things worth your time, ordered by leverage. From the home page if you want the full sprawl, or /principles if you came for the takes.

01

Subscribe to The Production Agent

Weekly Substack on shipping agentic AI to production: orchestration patterns, eval harnesses, governance, and the failure modes that don't make it into the launch posts.

Subscribe (free)
02

The pyAGI acquisition story

Published an early autonomous-agent Python package on PyPI in May 2023, before the agent wave. Acquired by Kyle Morris and Jeffrey for $10,000. Kyle Morris co-founded Banana.dev and is a founding member of AGI House. The artifact behind everything else on this site.

View on PyPI
03

agent-audit-kit - OWASP Agentic + MCP Top-10 SAST

Open-source security scanner for agentic AI systems. 375 rules across OWASP Agentic Top-10 (10/10) and OWASP MCP Top-10 (10/10). GitHub Marketplace-listed. Drop it into any CI pipeline.

Open microsite
04

Provael - red-team & assurance for physical AI

Open-source (Apache-2.0) harness that red-teams vision-language-action robot policies in simulation and reports an attack-success rate with a 95% CI + benign control. Authored the Embodied AI Security Top 10. Measured result (simulation only): under a roleplay instruction, a real SmolVLA policy left its safe envelope in 84% of matched pairs (42/50) vs a 2% benign control (1/50), 95% CI [62%, 100%]; fragility under an out-of-distribution instruction, not steering. Action-layer red-teaming: the embodied twin of my agent-security work.

Visit provael.com
05

Disclosure #001 - MCP STDIO Config Injection (CVE-2026-30623)

Inaugural entry in the agent-audit-kit Public Disclosures stream. Why MCP STDIO is RCE-by-design across 150M+ downloads, how AAK-STDIO-001 detects it, and what the remediation actually looks like.

Read disclosure
06

Principles - how I build production AI agents

13 opinionated rules from 3+ years shipping agentic systems. Audit-trail first · capability leases over keys · deny by default on tool calls · MCP STDIO is untrusted · vendor-multi by default · the patterns that survive the acquisitions.

Read the rules
07

VAJRA - physical AI build-in-public

Autonomous find-and-fetch robot on Jetson Orin Nano Super + ROS 2 Humble + NanoOWL (TensorRT-optimized OWL-ViT). Same governance instincts as Agentify (deny-by-default, budget routers, eval harnesses) re-applied to physical actions. Tabletop v0.1 milestone in motion.

Open VAJRA
08

Production work at Attri.ai

Tech Lead at Attri.ai. Agentify (15+ specialized agents in production), the Cascading Router cost strategy, E2B sandboxing, and the multi-cloud on-prem fork. Case studies for MannSetu (voice-first AI mental wellness) and Agentify (the platform).

Open case studies