[Build logs & notes]

Notes from the build loop.

> Agents, evals, scraping systems, robotics, and the bugs between the diagram and the running code.
Architecture, failure modes, benchmarks, and the occasional opinion.

[Published]8
// Latest post: 5 November 2026
AI systems / research / software
Editorial illustration of an agent loop calling fetch, retrieve and ask tools while an answer streams out with citations.
[Pinned]

My Biggest Engineering Work: Building a Harness From Scratch

A streaming agent loop built in a month with AI agents, and a 600-question eval.

[Read it]
Illustration of a home server routing polled Git changes to isolated test containers and a separate deployment gate.

GitHub for Review, a Home Server for CI/CD

Self-hosted CI/CD with polled refs, isolated PR tests, and exact-SHA deploys.

[> READ]
Token Bar illustration showing local Codex, OpenCode, and Claude Code usage flowing into a compact menu bar dashboard.

Counting Tokens Without Building Another Dashboard

A local macOS view of Codex, OpenCode, and Claude Code usage.

[> READ]
Agent memory write path from authenticated request through staged mutation to committed durable state

Agent Memory Is a Write-Path Problem

Bounded per-user agent memory with trusted identity and transactional writes.

[> READ]
AI chat analytics pipeline from live conversation to enrichment and reporting

AI Chat Analytics Is a Data Pipeline, Not a Prompt

Replayable chat analytics with idempotent writes and durable cursors.

[> READ]
Quarter Clock

Quarter Clock

A local Svelte app that shows whether yearly projects keep pace with time.

[> READ]
How I Scraped 500k Jobs Without Another Hosting Bill

How I Scraped 500k Jobs Without Another Hosting Bill

A single-host job scraper built with FastAPI, local processes, and Supabase.

[> READ]
Freud, Cultures, and My Own Ego

Freud, Cultures, and My Own Ego

What Freud's model can explain, and where my cultural analogy falls apart.

[> READ]