Project summary

When a fleet of robots fails in the field, the problem occurs once and then disappears. Engineers analyze logs, debate graphs, and ship a guess. I've written a runtime that replays a recorded failure exactly. There is no real-time clock or real-time input during replay, just the data from the recording, so the exact same failure occurs every single time. You can pause it, fork it, tweak a single variable, and observe a different result. Every failure becomes a permanently reproducible regression test.

Demo: https://drive.google.com/file/d/1l0HgizYSsFqqf0e-FrEep-Sm70kjAVIs/view?usp=sharing

This already exists for pure software (Antithesis, $105M raised) and for LLM evaluations (Inspect Evals and Equistamp, both funded on this platform). It does not exist yet for physical robots. And physical robots are increasingly controlled by trained policies which fail in unforeseen ways. An ex-Cruise validation lead told me they had 150-200 people working on verification and still could not reproduce 10-20 percent of field failures. Everyone smaller has nothing.

What are this project's goals? How will you achieve them?

First, deliver the first paid incident study. A robotics firm in Chile offered to pay $2,500 for one before I tried selling to them, and it happens this month. Second, make the ingest layer generic over other people's hardware. The engine works today and three real drones fly under it, but everyone logs differently, so each new study produces a reusable adapter. Third, the part relevant to this platform: release an open corpus of replayable fleet failures, so "can your stack reproduce this incident?" becomes an answerable question instead of a marketing slogan.

The theory of change is straightforward. You cannot do safety analysis on an incident that cannot be reproduced, and right now nobody can reproduce AI-driven robot failures. This is the same gap evals infrastructure filled for language models. I am not claiming this is alignment research. It is the tooling that makes safety analysis of these incidents possible at all, the way Inspect Evals does for model evals.

How will this funding be used?

All compute and hardware. About $10K of compute for large fault-injection runs, thousands of forked replays per incident. About $6K of test hardware: extra airframes, radios, and a ground robot to prove the engine works beyond drones. About $5K for adapters for the public corpus (ULog, MCAP, ROS 2 bags). About $4K for hosting the corpus. If a backer prefers equity, I'm happy to do a SAFE the way Equistamp and Seldon did.

Who is on your team? What's your track record on similar projects?

It's just me. 17, full time since June, based in San Francisco.

What works today: a deterministic engine with three physical drones flying under it. A simulation environment running four instances of real ArduPilot firmware, where killing a radio mid-flight triggers the failsafe at 4.9 seconds on every rerun. A real crash from last November replayed alongside a fork where the crash does not happen.

Before this I built a fleet orchestration system for search and rescue and cold-emailed 644 rescue organizations. 45 replied, none had budget before 2027, which taught me which market can actually pay. Earlier: a first-author ML paper in a peer-reviewed journal in high school, a Go engine that beat CrazyStone 16 of 20 games, and CTO of a 1517-backed drone startup at 16.

What are the most likely causes and outcomes if this project fails?

The most likely failure is that the adapters don't generalize. Every robotics stack logs differently, and if translating a new company's logs into replayable form stays expensive, this remains a service for a few fleets instead of infrastructure. The corpus stays small and the money buys limited results.

Second, timing. Fleets below the giants may not feel enough pain to pay until AI deployment scales further, and revenue stalls while the project lives on grant runway.

Third, I'm one person doing sales, delivery, and development, and one of those may get too little of my time. The part that will not fail is the determinism claim itself. It is built and enforced by tests.

How much money have you raised in the last 12 months, and from where?

$1,000 non-dilutive from ODF's Solo Grant program, committed and disbursing this month. A 1517 Fund Medici grant, $1K plus compute credits.