Solution architect manager by day, product builder by night. I ship agent systems and the gates that keep them honest.
Eighteen years at the systems layer, from network processors and IP/MPLS stacks to LLM agents. Today I lead an industry solutions team at a major cloud provider, building a media AI platform that agents call as tools, and I run a portfolio of AI products on my own time. One pattern connects both: let agents do the work, and make every gate deterministic. Evaluation sets, provenance states and human approvals that fail closed.
This page tries to show three things: hands-on depth (what shipped, with numbers), where I think agent tooling is heading (what I wrote), and product judgement (what I launched and how it is measured).
Launched in the window: NextPlace, BedFit Atlas, Memos 0.7.0 and Parts Lineage. Most of the merged changes are not features. They are gates: provenance checks, evaluation suites, publication rules and recovery paths that keep the pipelines dependable between releases.
focus: - Agent platforms: MCP tool gateways, multi-tenant runtimes, AgentCore - Evaluation & observability as release gates (Langfuse, golden sets, holdouts) - Autonomous engineering loops with deterministic verification - GEO/SEO growth engines that publish, measure and improve under review stack: languages: [Python, TypeScript, Swift, Go] cloud: [AWS (Bedrock, AgentCore, ECS, CDK), Cloudflare (Workers, D1, R2, Pages)] models: [Claude, GPT, Gemini, DeepSeek, MiniMax, Qwen] tools: [Claude Code, Codex, Langfuse, Playwright, Terraform] motto: "Let the agents build. Make the gates deterministic."
Three tracks, in the order I spend my time on them. Everything listed here shipped or changed in the last quarter.
The day job: a platform that agents consume as tools, and the evaluation machinery that decides what ships.
Tooling that lets a model plan, build and verify changes while the human only approves backlog and reviews evidence.
Established products with real users and a steady release rhythm. Each has its own ingest, publishing and measurement loop, and every change still passes tests and my review before it ships. I use them to test the ideas above against real traffic, not demos.
Built to learn a stack or to keep a promise to my kids. Still held to the same test gates.
Earlier in 2026: ShotAI (AI-native video asset management client), ScienceComics (interactive science comics), PolyTrader (prediction-market quant system).
The systems the three tracks grew out of: document intelligence for the legal industry, early agent workflows with their evaluation harnesses, and the GenAI MVPs shipped with customers at the cloud provider.
80 articles, 67 weekly newsletters and a 78-term glossary on this site, all written from systems I actually run. Recent pieces focus on one question: which parts of the agent stack are becoming infrastructure, and which are still choices.
Twenty link-only reports written for teams I work with, each with primary sources and hands-on tests. Shared on request.
Core contributor or maintainer of 10+ repositories under aws-samples and awslabs. I build tools I wish existed.
From low-level network processors to LLM agents, always at the systems layer.
Lead a team of forward-deployed engineers, SDEs and applied scientists delivering AI-native solutions to APAC and global customers. 1,000+ deployments, 100+ production launches, 10+ open-source repositories. Current focus: the media AI platform and evaluation gates above.
Built the cross-functional team and shipped LLM-powered document intelligence for personal-injury law: extraction with validation layers, RAG with citations, case association, and an evaluation framework for prompt and model iteration.
Built carrier-grade distributed systems. Hardware-software co-design, network processors, IP/MPLS protocol stacks, segment routing.
7 patents granted across US, EU, and China, spanning network resource optimization, SDN, packet processing, and distributed systems.
Happy to talk about agent platforms, evaluation gates, or any of the products above. Code first, slides second.