Notes from building the whole thing
AI engineering and modern architecture, written from production rather than from a blog post about production. Mostly what went wrong and what the fix turned out to be.
- 2 September 2026 · 6 min · agents, claude-code, reliability
What actually goes wrong with a team of agents
Running architect, engineer, QA and dev-lead agents across three production repos. The interesting failures were not bad code — they were agents manufacturing their own authority.
- 1 September 2026 · 6 min · architecture, reliability
When the API places real orders
A trading daemon runs unattended for six hours a day on a real brokerage account. The engineering that matters is not the strategy — it is what happens when something restarts at the wrong moment.
- 31 August 2026 · 7 min · evals, llm, cost
What a prompt change actually costs
Everyone asks whether the cheaper prompt hurt quality, and almost nobody answers with evidence. Most of the difficulty turns out to be statistical, not architectural.
- 30 August 2026 · 5 min · architecture, cost, caching
Bounding a billed API with the shape of its cache key
A public endpoint that calls a metered API is an unbounded bill. Sorting two values before they become a primary key turns cost from a function of traffic into a function of distinct questions.
- 28 August 2026 · 7 min · architecture, cost, llm
Free tiers are a design constraint, not a discount
What changes when zero infrastructure spend is a hard requirement — including running a consumer product's LLM layer entirely on free open-weight models, and the measurements that made it work.
- 24 August 2026 · 6 min · mcp, agents, architecture
Read-only by default: how mcp-multi-db enforces it
A walk through the two layers that make an MCP server safe to point at a real database — a SQL-text guard that is allowed to be imperfect, and a connection that cannot write.