Purpose-built
Every system is shaped around the work it has to do — not a template, not a demo, and not a wrapper around someone else's product.
NewSwarm extraction in openextract-ts
Production AI systems for teams and agents
Purpose-built for the difficult parts — extraction, agents, evaluation, and the infrastructure underneath. Built in the open. Owned by you.
The stack we build on
- TypeScript
- Python
- Rust
- Next.js
- AI SDK
- pydantic-ai
- Postgres
- MCP
- Zod
- Vercel
A different kind of AI partner. Small senior teams, shipping systems you keep — open wherever it can be.
Powered by models, not guesses
Multiple models read the same input and reconcile their answers. Confidence, provenance, and the source clause travel with every field.
Owned by you
The code, the models, the infrastructure, and the operating knowledge are yours from day one. No lock-in, no black boxes.
Selected systems
Useful AI, built all the way through
Four systems we build and maintain in the open. Nothing here is a mockup — read the source, run the benchmark, install the package yourself.
Open source / TypeScript01
openextract-ts
The TypeScript line of openextract. One call takes a file to a validated Zod object — and a swarm runs several models over the same document, then reconciles their answers by merge, vote, or first success. An MCP server, a CLI, a terminal UI, and a local web cookbook ship with it.
- Runtime
- Node 20+, TypeScript
- Swarm reduce
- merge, vote, or first
- Also ships
- MCP server, CLI, TUI, web UI
- Test coverage
- 100%, enforced in CI
- License
- MIT
Open source / Python02
openextract
Documents, images, audio, and video into a validated Pydantic model in one function call. Bring your own model — eleven providers work out of the box, and the same schema holds across all of them.
- Runtime
- Python 3.12+
- Providers
- 11, via pydantic-ai
- Test coverage
- 100%, enforced in CI
- License
- MIT
Applied system / TypeScript03
Open Lease Audit
A portfolio integrity console for commercial leases. Define the fields that matter, stream an abstraction of every document, then reconcile the whole portfolio for conflicting terms and contradictions.
- Per field
- Source clause + confidence
- Findings
- Ranked by severity
- Intake
- PDF, text, Markdown
- License
- MIT
Open source / Rust04
pdfmd
PDF to Markdown with nothing underneath it. The object-graph reader, the DEFLATE decoder, the font decoding, and the column and table heuristics all live in the crate.
- Dependencies
- 0
- Throughput
- ~3,900 pages/sec
- Test coverage
- 99%
- License
- MIT
How we work
Your systems. Your edge.
We build inside your environment, not a demo. Every decision is made to create durable capability rather than dependency.
Start a conversation- 01
Find the leverage
We map the system, pressure-test the opportunity, and define the smallest valuable release.
- 02
Build the real thing
Senior engineers work directly with your team and ship into your environment — not a sandbox.
- 03
Leave you stronger
The code, models, infrastructure, and operating knowledge are yours from day one.
Have a hard problem?
Let's build what should exist
Tell us where the work is slow, brittle, or impossible today. We'll respond with a clear point of view.
The Mellow Signal
A useful signal in the noise
Brief, opinionated notes on models, agents, infrastructure, and the work of putting AI into production.
No noise. Unsubscribe anytime.