We can't find the internet
Attempting to reconnect
Something went wrong!
Attempting to reconnect
Writing · Tag
12 posts on engineering. Or browse the full writing index →
A Tencent benchmark across 20 languages found Elixir at the top of LLM code-completion rates — Claude Opus 4 hit 80.3% on Elixir vs 74.9% on C#. The reasons aren't an accident; they're the same boring properties that have always made Elixir pleasant, now compounded by AI.
async/await and goroutines solve scheduling. The BEAM solves failure. Why most concurrency pain is actually failure-isolation pain — and only the actor model plus supervision trees fix it.
Open the repos behind the agent tooling you run — Ollama, the MCP SDKs, the orchestration engines — and it's all Go. Not because Go is good at AI. Because an agent tool is a concurrent network daemon that ships as one binary.
The "is Ruby dead" obituary runs every year. It confuses hype velocity with shipping velocity — and Rails 8 quietly deleted the actual reasons people left: Redis, Sidekiq, the Node build step.
Migrating an AI-first product from GCP to Azure cut $350K from infrastructure spend over six months. The negotiation that mattered more than the architecture, the $50K we accidentally cost ourselves back, and the four migrations I'd refuse to do today.
How we moved 225K+ users with $400M+ in fintech assets from AWS Cognito to Auth0 without forcing a password reset, breaking MFA, or interrupting active sessions. The lazy-migration pattern, the gotchas, and what I'd do differently.
Most engineers using Claude Code see a 10–15% speedup. The teams seeing 40–55% aren't typing faster — they're sequencing work differently. The four modes I use AI in, what to never delegate, and how to get a skeptical team across the line.
We started with ten Ruby and Elixir services serving real-time messaging for 450K students across 900+ universities. Two years later we had six, fully Elixir, and on-call alerts had halved. The migration order, the patterns we leaned on, and what I'd do differently today.
Most AI startups try to fix the model when the real problem is they can't see what the model is doing. The four-layer AI telemetry stack, the tooling to use, and how proper instrumentation cut a Lavender hallucination rate by 40% without touching the model itself.
The hardest part of agentic AI in 2026 isn't getting the agent to do the work. It's knowing when to override it. The four-level autonomy ladder, the five signals an agent is going off the rails, and a real example of catching one before it shipped a quietly broken auth flow.
A fractional engineering engagement starts with a codebase you've never seen. You have ninety minutes to form a useful POV before the kickoff call. The seven-step triage I run, the two questions I bring back to the founder, and how AI tooling has accelerated the process.
Most 'we delivered late' stories trace to one decision: the team scoped the first slice too big. The vertical-cut rule, the deploy-by-Friday filter, the pattern that breaks the heuristic, and a real before-and-after example.