What I'm building.

Drop a thought. Alter knows what to do with it.

• 11services
• 289Klines of TypeScript
• 40%of it is tests

Live, paid, and the thing I open first in the morning. I typed almost none of it.

You drop in whatever you're thinking; it works out what the thought is, files it where it belongs, and answers the only two questions that matter: what needs you today, and where everything else stands.

• TypeScript
• NestJS
• pgvector
• Redis
• Docker

An org of agents that builds the product, and the rules that stop it becoming a black box.

• 178merged PRs
• 9months
• 1human, with a day job

The hard part wasn't getting agents to write code. It was the rules that stop them colliding, marking their own homework, or talking their way past a limit.

One person, 178 merged pull requests in nine months. Most of the code was written by agents working in parallel, which only works because of what surrounds them: a sharded state model, independent verification, capacity limits that can't be talked around, and exactly three places a human has to say yes.

• Claude Code
• Multi-agent orchestration
• GitHub Actions
• Docker
• Bash

An AI filmmaking tool built around revision, not one-shot prompts: seven AI departments that argue their way to a locked film, on the record.

• 7departments, one artifact each
• 4gates, ideation to post
• 5roles in the one harness

It began as one agent and a dependency graph. It is now seven departments and one rule: a gate cannot close until every department the locks bind has read the latest version.

A tool for making short AI films the way films are actually made, by revision. Version one was a single agent revising a typed dependency graph with locks and cascades. The current version is a council of seven AI departments, each the single writer of one artifact, run by one harness through four gates over an append-only event log that every screen is a projection of.

• Claude API
• Multi-agent orchestration
• Event sourcing
• React
• Product design

One verification pattern, built once, that two document-intelligence systems run on.

• 1pattern, built once
• 2products running on it
• 4shared steps

Two products, different teams, different documents, different rules. The same program underneath.

Every check in a regulated document workflow has the same shape: gather the evidence, extract what it says, reconcile it against a system of record, and report a finding that cites its source. I built that shape once, as typed contracts, generic sub-agents and a read-only data layer that is an allow-list rather than a query builder, so that two very different products only had to supply their rules.

• Python
• Google ADK
• LiteLLM
• Pydantic
• Azure Container Apps

Every letter investigated and evidenced before a colleague opens it.

• 5parts to every case pack
• 2lanes, one graph
• 3engines in parallel

The colleague's job collapses from investigate to verify.

An asynchronous agentic system that reads incoming dispute and complaint correspondence, classifies it against a governed knowledge base, checks the account's documentation, verifies the correspondent, raises investigation flags, and assembles a case pack in which every conclusion links to its evidence. The colleague's job collapses from investigate to verify. Built end to end; in testing before pilot.

• Python
• Google ADK
• GPT vision
• Azure Document Intelligence
• Microsoft Fabric

Proving an account qualifies before it goes to legal, with evidence rather than sampling.

• 5checks per account
• 3verdicts: pass · fail · review
• 2review modes, never mixed

Every account, not a monthly sample. The reviewer's job becomes confirming exceptions.

Before an account is placed with a law firm, someone must confirm the required documents exist, are genuinely what they're indexed as, and agree with the system of record. Today that is a manual audit of a monthly sample. This system produces an evidenced verdict for every account (presence, authenticity, extracted fields, reconciliation), with every conclusion cited to a page. Most use cases built; template validation and package completeness remain.

• Python
• Google ADK
• GPT vision
• Microsoft Fabric
• Pydantic