The Most Dangerous Code Is the Code That Assumes
Every system has invariants—conditions that must always be true. The most dangerous ones are the implicit invariants that nobody writes down.
Articles
Long-form writing on software engineering, systems design, and building things.
AISpec: RSpec for AI Behavior, Not AI Outputs
LLM outputs vary, but system rules shouldn't. AISpec brings invariant contracts, confidence intervals, and Rails-native testing to your LLM pipeline.
Why AI Doesn't Replace Software Architecture — It Makes It More Valuable
AI can generate a thousand lines of code in a minute. It still can't tell you whether those lines belong in your system. As code becomes abundant, software architecture becomes the scarce resource.
The Two Sources of Noise Every LLM Evaluation System Must Handle
LLM evaluation is probabilistic. If you treat your LLM judges like unit tests, you'll chase ghosts. Here is how to separate the two sources of noise and build an evaluation pipeline you can trust.
The Outbox Pattern Isn't Free
The Outbox Pattern solves the dual-write problem, but it isn't free. From database write amplification to retention strategies and duplicate execution, here are the real costs of transactional reliability.
Why Performance Improvements Compound
Mature systems rarely waste time inside algorithms. They waste time moving data between them—and that is why performance fixes compound.
Stop Sending Embeddings as JSON: A Faster Binary Serialization Pattern for AI APIs
JSON is one of the most expensive ways to move embedding vectors. Here's how switching to binary Base64 float buffers cut payload size by 75% and reduced serialization latency by 98%.
More Threads, Less Throughput: When AI Servers Fight the Scheduler
When you need more throughput, adding threads seems obvious. But under the wrong conditions, concurrency can grind your system to a halt. Here is why your AI server is fighting the scheduler—and how to fix it.
Why Most Technical Migrations Fail (And How to Run Them)
Every senior engineer has lived through a migration that got 80% finished, stalled for a year, and left the team maintaining two versions of the same system forever. Migrations rarely fail because the new architecture is bad; they fail because of incorrect system boundaries, lack of dual-writing strategy, and ignoring human momentum.
Lessons from Leading a Production Incident (Triage under Pressure)
When the database CPU spikes to 99% or your background job queue has a backlog of 200,000 tasks, your engineering skills are only half the battle. Leadership during an incident is about command, communication, and systematic triage. Here is what I've learned from managing production fires.