The Build Log
I build to think. Most of these started as a problem in my own week and turned into a system, a tool, or a product. A working log — not a highlight reel.
A scored index of 1,600+ software categories on two axes: how much a category sets you apart, and how buildable it now is with AI. Every one lands on Build, Buy, Bridge, or Beware. Hybrid search, a public directory, and a paid Pro tier.
Supabase · pgvector · MCP · Astro View →
My personal operating system. One entity graph — projects, people, ideas, decisions — that captures from conversation, links itself, and surfaces what needs attention. Runs as an API with a memory layer that learns across sessions.
Python · SQLite · FastAPI · MCP
The pipeline behind my long-form writing: a ranked backlog, drafting in my own voice, a mandatory punch-up gate before anything ships, and branded imagery. Backlog to published, end to end.
Astro · Remotion · gated publish
The public face of the index: 1,600+ category pages and 169 domain hubs, built for human readers and AI answer engines alike. Structured data and one-line answers so the framework gets cited, not just visited.
Astro · structured data View →
A book-production app that takes a manuscript to launch-ready assets. I built it to ship my own book; it generalizes to any author.
Node · content pipeline
A unified cockpit that pulls my scattered tools, links, and automations onto one board. The place I actually start the day.
Supabase · React
The routing layer for working across AI models: which model gets which task, a contract format for handing off big builds, and a fresh-context review gate that checks the work before it's called done.
Multi-agent · review gates
A prompt-to-video factory. Describe a quote card or a data animation, get a rendered clip — driven by code, callable from any project.
Remotion · React
A writing filter that strips the tells of AI prose — inflated words, em-dash pileups, forced rules of three — so anything I publish reads like a person wrote it.
Style engine
Search that went from failing to excellent. An audit found the system was matching on keywords only and whiffing on plain-English questions. I rebuilt it as hybrid lexical + vector retrieval with embed-on-write — graded F to A, locked by regression tests.
pgvector · embeddings
Self-analytics for the whole system: every session, what it cost, what it produced. So the thing that runs my life can be measured, not just felt.
Python · analytics pipeline
An email assistant that reads an inbox, works out what each message needs, and drafts a plan — creating tasks and surfacing what matters, without ever hitting send. Runs headless on a schedule; a person approves every action.
Headless agent · planner / executor split
An AI chief-of-staff for a busy inbox: it sorts, ranks, and routes, then writes the results back to my dashboard so the day starts pre-triaged. Permission-contained — it reads and proposes, it doesn't act on its own.
Scheduled agent · write-back
Some of these are live and public; others run privately behind the scenes. I write about how they get built in The Build Report.
Bi-weekly analysis of software categories through the B4 Framework. What to build, what to buy, and how to use AI to make better decisions for your company.