( Writing )

Notes fromthe build.

Long-form engineering writing — mostly real-time voice AI: the latency, the pipelines and the systems behind making machines listen and talk. Fewer posts, more depth.

04 piecesNewest first

Cover — The New Bottleneck in Software Is Understanding

Latest

Nº 04Sep 15, 2026·8 min read·Craft

Perspective

The New Bottleneck in Software Is Understanding.

Execution used to be the expensive part of building software, and the pain of it quietly taught us how systems behave. That cost is gone — and so is the lesson. What that does to correctness, to hiring, and to anyone starting out now.

Read the essay

Earlier pieces

Nº 03Jul 6, 2026

Open Source Is Closing the Gap Faster Than Anyone Expected

Open-source coding models went from non-competitors to a genuine cost-performance decision in a couple of months. Where GLM-5.2, MiniMax M3, and Kimi K2.7 actually land against Opus, GPT-5.5, and Gemini — and why the gap closed faster than anyone predicted.

AI5 min read

↗

Nº 02May 18, 2026

Technical Implementation of a Real-Time Voice Agent

The hands-on build of a production voice agent — orchestrator code, audio plumbing, sentence buffering, barge-in, RAG, and the latency tricks that get one turn under ~850ms.

Voice AI18 min readPart 02

↗

Nº 01Feb 11, 2026

Anatomy of a Real-Time Voice Agent

An engineering deep-dive into the architecture behind modern voice AI — streaming STT/LLM/TTS pipelines, voice activity detection, latency masking, and the orchestrator that ties them together.

Voice AI15 min readPart 01

↗

More pieces in progress.