Thomas (Tom) Crockett · Software engineer

I build the systems
agents run in.

Runtimes. Sandboxes. Durable workflows.
And the evaluations that keep them honest.

Founder, Eumemic. Formerly NASA JPL and Stripe.

I also build the interfaces people use to operate complex systems, from Mars-rover science planning to healthcare tools.

Explore my work Resume PDF
NASA JPL / 11 years

Interfaces for daily Mars-rover operations.

Tuple Health / React + TypeScript

Healthcare interfaces and developer tooling.

Stripe / Applied LLMs

77% fewer statement-descriptor entry errors.

Selected work

01 — 04
01 / Agent infrastructureOpen source · Python

aios.

An agent runtime built
to keep going.

Tools run in the background. Sandboxes survive restarts. Workflows recover after a crash. An append-only event log records what happened.

I built aios as the infrastructure for Eumemic, a company expressed in code: workflows orchestrate agents, which orchestrate further workflows. The goal is zero-human operations. The system has merged hundreds of PRs, including changes to aios itself; higher-risk changes wait for my approval.

Explore aios

Inside the runtime

MessageTriggerWorkflow
01Durable job queueRe-enter the next agent step
02Agent stepReceive · reason · act
03Background toolsKeep the agent responsive
Append-only event logPostgres

Durable state. Recoverable execution.

  • Persistent sandboxes
  • Durable workflows
  • Tenant isolation
  • REST + MCP + SDK
02

AI Legion

Multi-agent systems · April 2023

Agents working together, from the first commit.

Released in April 2023. Agent-to-agent messaging, plug-in tools, a typed action language, and a memory manager that worked within a token budget. The starting point for the runtime work that became aios.

  • TypeScript
  • Message passing
  • Tool use
Explore the code
03

RagZoom

Hierarchical memory · Evaluation

Relevant passages verbatim. Everything else summarized.

A memory engine built on binary summary trees. Retrieval varies resolution with the query: relevant passages stay verbatim while the surrounding context is summarized under a token budget. I ran controlled experiments separating retrieval, summarizer, and reader effects; simpler retrieval performed better. The code and results are public.

  • Python
  • Retrieval
  • Controlled experiments
Explore the code
04

infinislop.ai

Generative video · Live product

A TV channel that writes its own next scene.

A director plans the story; a showrunner turns it into shots. Video generation, streaming, persistent story state, and usage-based funding come together in a continuous show. Built with Python, FastAPI, SQLite, and HLS.

  • Agent orchestration
  • Video pipelines
  • Streaming
Visit infinislop

Software in daily use.

Experience

Stripe

Jul 2022 – Jan 2025

Shipped an LLM feature for merchant onboarding that cut statement-descriptor entry errors by 77%. Built Beehive, a service that shields known-good users from incorrect action by fraud models.

Tuple Health

Oct 2016 – Mar 2021

Built healthcare web interfaces in TypeScript and React. Owned the company monorepo and developer platform: multi-project TypeScript, Webpack, ESLint, Jest, and GitHub Actions CI.

Storybook and Chromatic were part of that platform, alongside the application code and shared tooling.

NASA Jet Propulsion Laboratory

Jun 2005 – Oct 2016

Built Maestro and MSLICE, science-planning tools used in daily operations of the Mars rovers Spirit, Opportunity, and Curiosity.

Built Timely, a multiscale timeline visualization interface in TypeScript and D3. Core contributor to OnSight, the HoloLens interface for planning Curiosity activities on a virtual Mars. The OnSight team received NASA’s Software of the Year Award.

TypeScript & developer tools

Open source

Created runtypes, composable runtime validators that derive static TypeScript types. Built use-methods, a method-style alternative to React’s useReducer, and contributed TypeScript definitions for Material UI components.

Explore runtypes More TypeScript work