UAV · free detectors, Pro AI Detectors shipped · AI in development
Flight Log Analysis
Drop a flight log. Ask it questions. "Why did it crash?" "What was wrong with the tune?" "What should I change before the next flight?"
Stage one already exists, and it is free. The parsing and detection half of this product is published as flightdx under Apache-2.0 — ArduPilot dataflash and PX4 ULog parsing, seven fault detectors, and the measured value printed against the threshold for every finding. You can run it on a log right now, without a waitlist and without an account.
What is not built is the layer above it: the one that reads those findings and tells you which is the root cause. That distinction is not marketing caution — on a mistuned attitude loop the first alarm and the actual cause imply opposite actions, and getting it wrong is worse than staying quiet.
The idea
Flight logs already contain the answer — attitude, vibration, voltage sag, EKF innovations, mode changes, failsafes — but reading a dataflash log is a skill that takes years, and most crashes get diagnosed by forum thread. Flight Log Analysis pairs real log parsing (ArduPilot dataflash and ULog) with AI-assisted diagnosis grounded in the actual telemetry: it cites the log messages and timestamps behind every claim, suggests plausible causes ranked by evidence, and proposes concrete parameter or hardware fixes.
What asking it looks like
Ask it: "Quad dropped out of the sky about 40 seconds in — why?" — and it answers with cited telemetry:
VIBE.VibeZclimbing 12 → 58 m/s² with accelerometer clipping from t≈31s,RCOU.C3pinned at maximum while the other three outputs drop (the classic thrust-loss compensation signature),XKF4.SVvelocity variance growing after t≈39s. Diagnosis: physical damage or imbalance on motor 3's propeller — thrust loss, then saturation, then attitude loss. Then the concrete part: inspect prop 3, enable the harmonic notch filter, re-fly, and compare the two logs'VIBEtraces with the free Param Diff.
Where it stands, honestly
Stage one is built and published; stage two is not. The parsing and detection layer is flightdx, Apache-2.0, installable today — it reads dataflash and ULog, runs seven detectors, and prints the measured value against the threshold that was loaded on the aircraft, so every finding is checkable rather than trusted. The escalation gate that collapses a continuous fault into one advisory is sentinel-realtime, and it absorbed 96.8% of raw detections on a real flight.
What is unbuilt is the judgement: taking a set of findings and saying which one is the cause. That is genuinely hard rather than merely unfinished. On a mistuned attitude loop the symptom is advised 1.719 s before the cause, the two imply opposite actions, and a free deterministic rule that answers "whatever alarmed first" gets it wrong every time — which is exactly why the measurements for this layer are published, including the ones that went against it. The example above is illustrative of the target, not of current output.
It also builds on the free ArduPilot Param Diff tool and its parameter-metadata dataset (~5,700 parameters per vehicle from ArduPilot's official definitions). Waitlist feedback shapes what gets built first: which vehicles, which failure modes, which log formats.
Join the waitlist
One email when it opens — founding waitlist members get launch pricing and a say in the first supported failure modes.
Flying today? The free ArduPilot Param Diff already catches configuration drift between flights.