Greg Little

In my 2011 MIT thesis I wrote: “I see programs that are able to repair themselves … hiring other workers to verify the solution before integrating the changes.” Then the workers were people on Mechanical Turk. Now they are AI agents, and I build them for production, along with the checks that say when to trust them. I also design real-time sync algorithms and protocols (CRDTs, Braid).

Experience

• Merged 30 PRs across 7 repos in 11 weeks, mostly written with coding agents and reviewed blind by two models, then by a person.
• Matched each desk’s prompt-cache lifetime to its traffic, cutting the sales desk’s cost per agent run by 16% on day one. The cost meter would have reported the savings as a 42% increase, so I fixed its pricing too.
• Built a “bug machine”: agents pick Jira bugs whose fix is certain and harmless, write a failing test, fix the bug and pass blind review by a second model. Each PR is a few lines plus its test, at $50–60 in model cost.
• Built a replay harness with blind LLM judges and tested it on 30 real agent runs: for the sales desk, a fresh context per run cost 20% less than a long-lived session, with no increase in errors. The judges were noisy, so I reran outliers before trusting the error counts. An early 45–70% savings estimate did not hold up. The harness also caught two bugs in agent skills, which I fixed (duplicate follow-ups: 6/6 runs → 0/6; drifting to other customers: 5/9 runs → 0/8).
• Traced 6 of 8 escalations blamed on staff to gaps in the written quote, mostly promises made on the estimate call and lost in drafting.

• Braid / Invisible College (braid.org): co-designed Antimatter, the first peer-to-peer sync algorithm that prunes CRDT/OT history under arbitrary partitions, down to zero overhead in steady state. Designed the Sync9 text CRDT. Co-authored the IETF Internet-Draft Braid-HTTP.
• Sole engineer on a stealth Figma-style collaborative design tool (2019–2024): full stack, with AI integration.
• Enso (2022–2024; CTO from 2023): end-to-end-encrypted real-time collaboration for text and files.
• SIV (2020): an implementation of Neff’s verifiable secret shuffle for e-voting.
• Other contracts: a traffic simulation for a personal-rapid-transit client; a GPU Equihash miner.
• Co-instructor, Bradfield School of Computer Science, San Francisco (2018): Languages, Compilers and Interpreters.

• Researched distributed hiring and on-the-fly teams; ran public experiments, e.g., hiring experts in 20-minute chunks.

• Shipped Operation: Blockade and Operation: Steel Tide (Infogrames, Windows, 2002). One of two programmers: built the 3D graphics, the game AI (aircraft movement, strafing-path planning) and the game editor.

Education

MIT: Ph.D. Computer Science 2011 (Programming with Human Computation, advisor Rob Miller); S.M. 2007; three TA appointments. Arizona State: B.S. Computer Science.

Selected research (first author; ~5,400 citations overall)

• TurKit (UIST 2010): crash-and-rerun programming, which reruns a script without paying again for human steps already done. An early form of today’s durable agent workflows. Covered by MIT Technology Review.
• Keyword Programming in Java (ASE 2007; Automated Software Engineering 2009): loose keywords → valid code.
• Koala (CHI 2007, IBM Almaden): web automation that became IBM CoScripter.

Skills

Python/Django · TypeScript/Node · Postgres, BigQuery · Claude Code and agent SDKs, prompt caching, LLM-judge evals, cost accounting · CRDTs and sync protocols · WebGL