Engineered by hand.
Run by agents.
Custom AI agents, the platforms they run on, and the operating models behind them. Built for a small group of companies that operate.
Agents that do the work, and everything they need to keep doing it.
One team owns the agent, the software around it, and the numbers that say whether it is working.
Tool use over typed schemas
Multi-step orchestration with idempotent steps, retries with backoff, and human-in-the-loop checkpoints where the stakes call for one. Retrieval over your own systems, not a generic index.
The software around the agent
Web properties, serverless functions, integrations and internal tools, built as versioned packs and deployed on a global CDN.
Model-agnostic by design
Frontier models selected per task and evaluated on your data. A swap is a config change, not a rewrite.
Numbers that tie out
Dashboards, mobile tools and financial workbooks that reconcile to the ledger, built for the people who open them daily.
We stay on after launch
Updates, new capabilities and the occasional rethink, on a simple retainer with a stamp on every release.
How it is built.
The stack and the discipline, so an engineer can read this page and know what to expect.
- Models
- Frontier models from every major lab, selected per task and evaluated on your data: Anthropic, OpenAI, Google DeepMind, Meta, Mistral. Prompt caching, streaming, structured outputs, and eval suites that run on every prompt change.
- Agents
- Tool use and function calling over typed schemas. Multi-step orchestration with idempotent steps, retries with backoff, and human-in-the-loop checkpoints. Regression tests on agent behavior before anything ships.
- Stack
- TypeScript end to end. Postgres with row-level security. Serverless and edge functions. Event-driven integrations with idempotent webhooks and dead-letter queues. Infrastructure as code. Static-first delivery with immutable, content-hashed assets.
- Delivery
- Trunk-based development. Continuous integration with a preview deploy per change. Versioned build stamps on every release. Automated visual, accessibility (WCAG 2.2 AA) and performance regression. Zero-downtime releases with one-step rollback.
- Observability
- OpenTelemetry traces, structured logs, error budgets, and per-client dashboards for uptime, latency and model spend.
- People
- Senior engineers with experience at major software companies, working directly with founders and operators. Small team by design.
Built for regulated operators.
Healthcare was the first client, so the controls came first.
The books are closed.
We take on a small number of companies and stay with them. If you have a project and want to know whether an opening comes up, write to us. You will get a straight answer either way.