The Assistant Answered. The Memory Never Arrived.
Eighty-four answered exchanges were declined by an agent memory integration, while the monitoring check missed the failure log. How we changed exchange identity and verified a live save.
Writing
17 pieces on applied AI systems (routing, evals, voice, inference and agents), each tied to work I actually built or ran.
Give Your Coding Agent a Maintained Map
A small, reviewable knowledge vault can give a coding agent the right architecture and decisions without asking it to trust stale notes over the code.
It Was Already in Your Data
The enzyme had been seen before. The pattern around it had not.
Lab Note: I Published My Request Router. Here's How to Train Your Own.
The final training run took about 24 seconds on our RTX 5090, excluding Python startup. Deciding what the model should learn, labelling 2,334 examples, and building a test that could reject it took the rest of the afternoon.
Train Your Own Request Router: Fine-Tuning ModernBERT to Pick the Lane in 45 ms
Every request to an AI system starts with a quiet decision: what kind of answer does this need? Words? A picture? A search? A voice message? A video?
We Trained Our Own Router. It Runs on a Jetson.
Before an AI system can answer, it often has to make a smaller decision: What kind of answer does this person need?
Your Employees Got Faster. Your Company Didn't.
McKinsey's 2026 global survey found that 80% of respondents said AI had improved their individual productivity.
Your Alert Is Only as Stable as Its Least Stable Label
At 4:17 a.m., our monitoring said one of our AI agents had stopped running its scheduled heartbeat.
Your AI Agents Are Building an Org Chart Whether You Realize It or Not
Give one AI agent a task and you have an assistant.
When AI Becomes the Junior Engineer, How Does the Junior Engineer Become Senior?
Nobody hires a junior engineer today and says, "You can't use AI. Learn everything by hand while the rest of the team uses it."
AI Inference Is Becoming a Supply Chain
Your inference dashboard says the model is fast. Your users say the product is slow. Both are telling the truth, and the gap between them is where AI infrastructure is changing fastest.
You Added the Safe Path. You Didn't Remove the Dangerous One.
We documented the safe way to do it. The dangerous way still worked. Guess which one we used twice in one evening.
Measuring OpenAI Compatibility for LangGraph Endpoints
An OpenAI-compatible endpoint is easy to demo and surprisingly hard to specify.
Stop Picking a Winning Model. Build a Winning System.
An HTTP 400 changed how I think about local AI.
Green Text Is Not a Gate: Four Ways Verification Becomes Theater
An AI agent can make a test suite green before a team has decided what green is allowed to mean.
From Barge-In to Floor Control: Engineering Voice Interruptions Under Uncertainty
Most voice-agent stacks begin with the same loop: transcribe the caller, send text to a model, synthesize the answer, and interrupt the synthesis when new speech arrives.
Full-Duplex Voice Is More Than Simultaneous Speech. It’s Floor Control Under Uncertainty.
The easiest voice-AI demo to understand is interruption.