Skip to content

Demo film

Full tour (20 min)

Every page in the product, narrated, with captions burned into the picture and twelve chapter marks. Seventy five scenes covering seventy two of the seventy three routes in the web application; the one it skips is the sign-in page.

Every number spoken in it is read from the live database at record time and substituted into the script, so the voice cannot disagree with the screen. That includes the unflattering ones: the tour says out loud that 788 of 598,179 objects carry a human verdict, that 30,863 of 30,865 reviews have no timing, and that the multi-camera benefit is measured over six sessions and is therefore small.

It was filmed on four recordings ingested for it: three dashcam clips from Bengaluru and KITTI drive 2011_09_26_drive_0005, whose import gave the corpus its first measured ego poses (154) and its first laser points (154 clouds, 18 million points).

Watch on YouTube · download the file (48 MB, H.264 1080p) · transcript

The captions are part of the picture, so there is no subtitle track to switch on and nothing to line up. The transcript is the same text as a separate file, kept in its own directory because a player auto-loads a .srt sitting beside a video and would draw it over the captions already there.

How it was made, and the two defects making it uncovered: the recorded tour.

Verification pass (4 min)

A record of a checking pass over the system rather than a description of it: more data downloaded and ingested, every readable route called against the live database, the HD map taken end to end, and the one route that failed. Every number spoken came from one of those runs, including the two results that are not yet trustworthy.

Download (21 MB) · transcript · the written report

The earlier films

Everything in these films is the real application, recorded live from a browser driving the running system: the real corpus (578,436 objects, 41,752 frames, 377 sessions), the operator's own DASHCAM-01 fleet footage in the editor, real GPU SAM segmentation, a real drift scan hitting the governance plane. Nothing is mocked and nothing is staged; the two labels the demo tools created were deleted right after filming, and the settlement and control-sample queues were never touched.

Narrated tour (4 min)

Voiceover, the full platform, and every canvas tool operated on camera: box, polygon, polyline, point-SAM, box-SAM, magic wand, measure, lane-aligned 3D cuboid, pose keypoints, adverse-region flag, mask brush, eraser, superpixel cells, and the Lanes / Semantic / Events mode rail.

Download the narrated film (16 MB, H.264)

Silent cut (3 min)

The same hands-on tour without narration - captions only.

Download the silent film (10 MB, H.264)

What the films show, in order

  1. The platform launcher, and the risk-ranked review queue with bulk accept / reject / relabel.
  2. The frame editor on real Bengaluru dashcam footage: every drawing tool, the AI family on the local GPU, measurement, cuboids, pose, adverse regions, mask surgery, and the mode rail.
  3. A 90-frame track timeline where one relabel fixes every frame at once.
  4. A settlement lot's acceptance sample - the human verdicts that decide whether 9,586 labels settle.
  5. The control queue that audits the auto-accept gate itself.
  6. Natural-language search, scenario discovery, driving events, corpus analytics, gold-set health, the training factory, jobs, sealed dataset delivery, and multi-format import.
  7. The governance console (a real drift scan, clicked on camera) and the autonomy console: a live daemon heartbeat, the per-class permission ladder, and the evidence behind every rung.

Two real defects were found by filming and fixed on camera: a daemon heartbeat that read as dead during its busiest tick, and a drift-recovery path that silently re-armed autonomous promotion. Both fixes shipped with regression tests proven by reverting them.