05 Writing
Thoughts on AI systems, engineering leadership, and building production software.
Put the Language Model Last
Cloudflare shipped Clef this week, so I built a voice agent where a decision model reads the room and the LLM only writes the words. What worked, what didn't, and what it says about how voice agents should be built.
Agent Frameworks Treat the Human as the Bottleneck
The agent frameworks are racing toward autonomy while the maintainability data goes the other way. Why I built Walter as a thinking partner instead of a spec pipeline, and what using it this year taught me.
Demystifying Claude Code: Beyond the Chat Interface
CLAUDE.md, skills, MCP servers, subagents, hooks, and agent teams — how the pieces actually work, when to use each one, and the patterns that emerge when they compose.
What I Learned Scaling a Team from 20 to 130 Engineers
From negative $50K a month to 130 engineers across 40 countries during a pandemic. The strategy that worked, the breaking points, and what happened after I left.
Building a Real-Time Voice Pipeline From Scratch
From a two-process prototype to production WebSocket voice. VAD tuning, Whisper padding tricks, and local-first inference with smart-whisper and Kokoro-82M.
GPU Inference Without the Cloud Tax
SPOT L4 instances at $0.28/hr vs managed services at $0.80/hr. Terraform automation, zero-trust networking, and container patterns for self-managed GPU inference.
What No One Tells You About the IC-to-CTO Pipeline
The standard IC-to-CTO advice is five recycled points stripped of context. Two stints, one scaling a failing org and one navigating a political minefield, taught me what the playbook leaves out.
Building Secure MCP Servers for Production
A practical guide to the architecture decisions, implementation patterns, and operational practices that separate production-grade MCP servers from the kind that end up in breach timelines.