Teng Li
ML Engineer @ Zoom · Seattle
I build AI agent systems for production — and write about the engineering underneath. Notes on ML systems, agent tooling, and a career spent closing the gap between what works in a demo and what survives real users.
Latest build · Hacktoberfest 2026
SlothPulse: your standup draft, with receipts
Git commits and an optional error trace become a Markdown draft through local Gemma. Follow each source reference, review the wording, and keep the note on your machine.
What I write about
ML & Engineering
Building and running ML systems in production — model serving, data pipelines, and the unglamorous infrastructure that keeps things working.
AI Agents & MCP
Field notes from shipping agents to real users: tool design, connectors, MCP servers, eval harnesses, and the gap between spec-compliant and actually usable.
Career & Industry
The longer view — how ML engineering careers compound, what's changing in the industry, and lessons from the inside of a fast-moving field.
Recent writing
All posts →-
My local AI summary had valid sources. It still connected the wrong facts.
AgentsEngineering
Building SlothPulse exposed three different checks: valid JSON, complete source coverage, and the right relationship between a blocker and its follow-up.
-
Instinct moves $1B a year through SMS and email. Its bottleneck is the one I've been measuring.
AgentsMCPEngineering
Instinct's growth curve isn't a model quality curve. It's a permissions curve. That lines up uncomfortably well with what I found scanning 36 MCP servers.
-
My eval said a perfect MCP server was broken. It was the eval that was lying.
AgentsMCPEngineering
Three rounds of calibration, four real servers, $0.60 in API costs — how I made an LLM-powered tool-selection benchmark fair enough to publish, and what it revealed about static lint rules predicting live model behavior.
-
I lint-scanned 36 popular MCP servers. A third of them are failing your agent.
AgentsMCPEngineering
Your MCP server can be 100% spec-compliant and still be unusable by an agent. I built mcpgrade to measure the difference — then pointed it at 36 popular servers.
-
Designing tools LLMs can actually use: five failure modes I keep seeing
AgentsEngineering
Your tool API can be perfectly engineered and still confuse every model that touches it. Field notes on why, from someone who integrates agent tools for a living.
-
Why this blog exists
Industry
Production AI agents fail in ways that never show up in demos. Someone should write those down.
Projects
SlothPulse
Local Gemma standup drafts with Git and error receipts. Preview the context, validate source bindings, and keep the Markdown on your machine.
mcpgrade
Lighthouse for MCP servers — scores whether an MCP server is actually usable by an agent, not just spec-compliant. 24 static rules + an LLM-powered eval, benchmarked across 36 popular servers.
