1 00:00:00,000 --> 00:00:05,831 This is LabeloxAV, a data engine for autonomous driving built around Indian roads. 2 00:00:05,831 --> 00:00:15,622 Everything you are about to see runs against a live database with 381 driving sessions, 42,766 frames and 598,179 labelled objects in it. 3 00:00:15,622 --> 00:00:18,142 No screen in this tour is a mock up. 4 00:00:18,142 --> 00:00:26,133 Where a number is small or a page is empty, that is the real state of the system and the narration will say so. 5 00:00:26,133 --> 00:00:29,265 The product is ten platforms over a single data spine. 6 00:00:29,265 --> 00:00:39,607 Annotation, review, measurement, governance, edge deployment and the rest each get their own surface, but they all read and write the same sessions, frames, objects and tracks. 7 00:00:39,607 --> 00:00:44,866 That is why a correction made in the editor changes what the gate measures an hour later. 8 00:00:44,866 --> 00:00:48,968 Work is scoped by project, and each project belongs to a domain pack. 9 00:00:48,968 --> 00:00:56,389 The autonomous driving pack brings the road ontology, the camera and LiDAR sensor model and the traffic scenario vocabulary. 10 00:00:56,389 --> 00:01:01,999 A second pack for physical security shares the same spine and none of the road specific code. 11 00:01:01,999 --> 00:01:03,015 Data arrives here. 12 00:01:03,015 --> 00:01:11,922 The importer takes dashcam video, ROS bags, MCAP recordings and raw sensor drives, and turns each one into a session with frames at a fixed timestamp. 13 00:01:11,922 --> 00:01:18,497 For this tour I ingested four real recordings, of which three are dashcam clips from the owner of this machine. 14 00:01:18,497 --> 00:01:23,399 The last is a KITTI drive from Karlsruhe with a sixty four beam LiDAR on the roof. 15 00:01:23,399 --> 00:01:26,190 Most teams already have labels somewhere else. 16 00:01:26,190 --> 00:01:36,672 The migration path reads COCO, YOLO, Pascal and CVAT exports, maps their class names onto this ontology, and records the mapping so that nothing silently changes meaning. 17 00:01:36,672 --> 00:01:40,332 Classes it cannot map are held back rather than guessed at. 18 00:01:40,332 --> 00:01:42,813 Every ingested session lands in the inspector. 19 00:01:42,813 --> 00:01:49,869 This is a Foxglove class viewer built into the product, so you can check a recording before you spend any labelling effort on it. 20 00:01:49,869 --> 00:01:52,350 The health gate at the top is the useful part. 21 00:01:52,350 --> 00:02:00,399 It refuses a session with broken timestamps, missing calibration or a camera that dropped out, instead of letting the problem reach the labellers. 22 00:02:00,399 --> 00:02:01,863 This is the KITTI drive. 23 00:02:01,863 --> 00:02:07,401 The timeline shows every topic on one clock, the camera alongside the LiDAR and the GPS. 24 00:02:07,401 --> 00:02:15,866 The panels are native, not an embedded iframe, so a click on the timeline moves the point cloud, the image and the ego pose together. 25 00:02:15,866 --> 00:02:19,035 Calibration is what makes three dimensional work possible. 26 00:02:19,035 --> 00:02:28,040 This session carries real intrinsics from the KITTI rig, a focal length of seven hundred and twenty one pixels over a twelve forty two by three seventy five image. 27 00:02:28,040 --> 00:02:33,266 Without this the system can still draw boxes, but it cannot tell you how far away anything is. 28 00:02:33,266 --> 00:02:35,102 Here is the vehicle's own motion. 29 00:02:35,102 --> 00:02:37,570 Speed, yaw rate and the trajectory it drove. 30 00:02:37,570 --> 00:02:42,046 The KITTI import brought 154 measured ego poses from the GPS and inertial unit. 31 00:02:42,046 --> 00:02:53,466 Before this import the entire corpus had three satellite fixes across forty one thousand frames, which is the honest reason most of the three dimensional work here was estimated rather than measured. 32 00:02:53,466 --> 00:02:57,044 A dataset is a frozen selection of frames with a content hash. 33 00:02:57,044 --> 00:03:07,133 Two people training on the same dataset identifier are training on exactly the same images and labels, which is the only way a comparison between two models means anything. 34 00:03:07,133 --> 00:03:08,866 Labelling starts automatically. 35 00:03:08,866 --> 00:03:15,219 You choose a session and the classes you care about, and the auto label runner works through it batch by batch. 36 00:03:15,219 --> 00:03:19,319 Batch by batch is a deliberate constraint, not an implementation detail. 37 00:03:19,319 --> 00:03:28,733 The runner holds a GPU slot, checks free video memory between batches, and stops if a training job wants the card, so a labelling run cannot take the machine down. 38 00:03:28,733 --> 00:03:33,622 This is the annotation canvas with real machine predictions on a real frame. 39 00:03:33,622 --> 00:03:35,708 The tools run down the left side. 40 00:03:35,708 --> 00:03:41,379 Box, polygon, segment anything, lane, drivable area, keypoints, amodal extent and depth. 41 00:03:41,379 --> 00:03:49,266 Every tool is registered in one place, so the keyboard shortcut you see in the panel is the shortcut that actually fires. 42 00:03:49,266 --> 00:03:51,924 The describe tool is the one worth stopping on. 43 00:03:51,924 --> 00:04:00,766 You type what you want in plain words, for example a cow standing on the median, and an open vocabulary detector proposes it with a segment anything mask. 44 00:04:00,766 --> 00:04:09,666 This exists because the ontology will never have a word for everything on an Indian road, and the alternative is that those objects go unlabelled forever. 45 00:04:09,666 --> 00:04:13,580 Lanes and drivable surface are annotated as geometry rather than boxes. 46 00:04:13,580 --> 00:04:18,222 The drivable mask now covers ninety nine point nine percent of frames in the corpus. 47 00:04:18,222 --> 00:04:22,248 It used to cover under four percent, and the gap was not a model problem. 48 00:04:22,248 --> 00:04:26,666 Three separate consumers were reading the masks in a format nobody had written. 49 00:04:26,666 --> 00:04:30,942 Attributes are swept across many objects at once instead of one at a time. 50 00:04:30,942 --> 00:04:36,975 Occlusion level, truncation, whether a rider is wearing a helmet, whether a vehicle is parked or moving. 51 00:04:36,975 --> 00:04:44,999 One pass over a hundred crops of the same class is far faster than a hundred visits to the editor, and it keeps the judgement consistent. 52 00:04:44,999 --> 00:04:47,323 Road furniture gets its own surface. 53 00:04:47,323 --> 00:04:50,244 Signs, signals, poles, barriers and markings. 54 00:04:50,244 --> 00:05:00,799 Signs carry an IRC sixty seven code, which is the Indian standard, so a mandatory sign and a cautionary sign are distinguishable without reading the pictogram. 55 00:05:00,799 --> 00:05:03,099 Behind the auto labelling sits an ensemble. 56 00:05:03,099 --> 00:05:07,150 Several detectors run over the same frame and their outputs are reconciled. 57 00:05:07,150 --> 00:05:10,107 Where they agree the label is cheap and probably right. 58 00:05:10,107 --> 00:05:16,732 Where they disagree is exactly where a human should be looking, and that disagreement is mined rather than averaged away. 59 00:05:16,732 --> 00:05:20,374 Agreement between models becomes pseudo ground truth for training. 60 00:05:20,374 --> 00:05:23,344 One lesson from building this is recorded on the page. 61 00:05:23,344 --> 00:05:28,500 Choosing the teacher pair by how many predictions each model made gave nine consensus labels. 62 00:05:28,500 --> 00:05:33,599 Choosing by how many frames the two models shared gave one thousand one hundred and ninety. 63 00:05:33,599 --> 00:05:37,284 Machine labels are proposals until a person rules on them. 64 00:05:37,284 --> 00:05:46,915 The queue orders work by how much a verdict is worth, combining model uncertainty, class rarity and whether a gate is currently blocked on that class. 65 00:05:46,915 --> 00:05:54,866 There are 528,793 objects waiting in review right now, which tells you honestly how far ahead of the humans the machine is. 66 00:05:54,866 --> 00:05:57,732 Grid review shows many crops of one class at once. 67 00:05:57,732 --> 00:06:06,331 The eye is very good at spotting the odd one out in a wall of similar images, and very bad at staying calibrated across a thousand single decisions. 68 00:06:06,331 --> 00:06:10,133 Accept or reject applies to everything selected in one keystroke. 69 00:06:10,133 --> 00:06:13,589 Rapid review is the keyboard only mode for a long session. 70 00:06:13,589 --> 00:06:15,287 One crop, one key, next crop. 71 00:06:15,287 --> 00:06:23,533 It records how long each verdict took, because the cost of a label in minutes is the number the whole economics of this system rests on. 72 00:06:23,533 --> 00:06:29,280 This is the full annotation table, filterable by class, state, source, session and reviewer. 73 00:06:29,280 --> 00:06:31,112 State is the important column. 74 00:06:31,112 --> 00:06:37,238 A label is in review, auto accepted, settled or accepted, and accepted means a person ruled on it. 75 00:06:37,238 --> 00:06:39,512 The machine can never write accepted. 76 00:06:39,512 --> 00:06:45,133 That is enforced in code and there is a test that fails the build if any code path tries. 77 00:06:45,133 --> 00:06:48,084 Any single label opens to its full history. 78 00:06:48,084 --> 00:06:58,203 Which model proposed it, at what confidence, which human touched it and when, what it was before each change, and which training run consumed it. 79 00:06:58,203 --> 00:07:03,333 You can answer where did this label come from for any of 598,179 objects. 80 00:07:03,333 --> 00:07:05,289 Quality is measured, not asserted. 81 00:07:05,289 --> 00:07:15,012 Precision comes from stratified samples with Wilson confidence intervals, so a class with twelve judged crops reports a wide interval rather than a confident number. 82 00:07:15,012 --> 00:07:17,147 Unmeasured is reported as unmeasured. 83 00:07:17,147 --> 00:07:19,933 It is never reported as zero and never as fine. 84 00:07:19,933 --> 00:07:22,077 Errors are typed rather than counted. 85 00:07:22,077 --> 00:07:33,866 A missed detection, a duplicate, a class confusion and a loose box are four different problems with four different fixes, and lumping them into one accuracy number hides which one you actually have. 86 00:07:33,866 --> 00:07:35,631 This page measures the reviewers. 87 00:07:35,631 --> 00:07:43,465 Control samples with a known answer are mixed into the queue, so agreement between people can be computed without anybody re-reviewing anybody. 88 00:07:43,465 --> 00:07:45,286 A finding worth admitting is here. 89 00:07:45,286 --> 00:07:54,333 Thirty thousand eight hundred and sixty three of thirty thousand eight hundred and sixty five reviews carry no timing, because the field was only wired up recently. 90 00:07:54,333 --> 00:07:59,440 An object across frames is a track, and a track is the unit that matters for driving. 91 00:07:59,440 --> 00:08:01,324 The corpus holds 13,807 of them. 92 00:08:01,324 --> 00:08:06,006 Eighty four percent of tracks used to blink out and restart as a new identity. 93 00:08:06,006 --> 00:08:13,666 Interpolation across gaps had been written but never run, and turning it on filled one hundred and thirty seven thousand gaps. 94 00:08:13,666 --> 00:08:16,124 Corrections apply to the track, not the frame. 95 00:08:16,124 --> 00:08:24,152 Relabelling a motorcycle as a rider used to reach one frame out of ninety three, because the editor and the review queue took different write paths. 96 00:08:24,152 --> 00:08:27,866 They now share one, and the propagation is visible on this timeline. 97 00:08:27,866 --> 00:08:31,790 A vehicle with several cameras sees the same object more than once. 98 00:08:31,790 --> 00:08:38,032 With calibration, a box drawn in one view projects into every other view in the rig, so you annotate once. 99 00:08:38,032 --> 00:08:44,333 The corpus has six multi camera sessions, so the measured benefit here is small and I will not inflate it. 100 00:08:44,333 --> 00:08:47,826 Between two labelled frames the vehicle moved a known distance. 101 00:08:47,826 --> 00:08:57,066 That motion alone predicts where a static object should appear next, without running any model, and it is enough to carry a label forward through a stretch of road. 102 00:08:57,066 --> 00:09:04,828 When a track breaks behind a bus and reappears, the recovery pass tries to rejoin the two halves using appearance and predicted position. 103 00:09:04,828 --> 00:09:11,733 A rejoined track is marked as rejoined rather than silently merged, because a wrong merge is worse than two short tracks. 104 00:09:11,733 --> 00:09:14,123 This is where the tour gets its real LiDAR. 105 00:09:14,123 --> 00:09:19,642 The KITTI drive brought 154 point clouds from a sixty four beam Velodyne, about 18 million points. 106 00:09:19,642 --> 00:09:29,200 Before this import the corpus had four hundred and fifty three thousand points in total, all of them estimated from single camera depth rather than measured by a laser. 107 00:09:29,200 --> 00:09:33,145 Cuboids are drawn in the point cloud with the camera image beside them. 108 00:09:33,145 --> 00:09:40,133 The ego trajectory runs underneath, so a box can be placed in the world frame and stays put while the vehicle moves past it. 109 00:09:40,133 --> 00:09:43,083 The linked view keeps the image and the cloud in step. 110 00:09:43,083 --> 00:09:48,761 Click a two dimensional box and the matching three dimensional box highlights, and the other way round. 111 00:09:48,761 --> 00:09:51,266 This is also how the lifted boxes are checked. 112 00:09:51,266 --> 00:09:57,333 A cuboid estimated from camera depth alone is compared against the laser points that actually fall inside it. 113 00:09:57,333 --> 00:09:59,577 Sessions on a map, by route and by city. 114 00:09:59,577 --> 00:10:07,000 The locations here are aggregated into cells rather than drawn as raw points, and I will come back to why in the privacy chapter. 115 00:10:07,000 --> 00:10:08,999 Verdyx is where models are measured. 116 00:10:08,999 --> 00:10:15,795 Accuracy is reported per class with an interval, and against a frozen gold set that no training run is allowed to touch. 117 00:10:15,795 --> 00:10:19,336 There are 15 registered models, each with its lineage recorded. 118 00:10:19,336 --> 00:10:24,133 What it was trained from, which teacher it distilled, and which dataset hash it saw. 119 00:10:24,133 --> 00:10:26,801 An average number hides the cases that matter. 120 00:10:26,801 --> 00:10:31,664 Performance is sliced by night, rain, occlusion and by the safety critical classes. 121 00:10:31,664 --> 00:10:33,798 Frames are also perturbed on purpose. 122 00:10:33,798 --> 00:10:39,491 Occlusion, dusk, rain, fog and motion blur are applied to real images and the model is re-scored. 123 00:10:39,491 --> 00:10:48,800 The current champion loses twenty four points of motorcycle recall under occlusion, which is the kind of thing an average accuracy score will never tell you. 124 00:10:48,800 --> 00:10:52,525 A scorecard is the summary a person actually reads before a release. 125 00:10:52,525 --> 00:10:58,475 Per class recall and precision, the slices that regressed, the gate result and the reason if it was blocked. 126 00:10:58,475 --> 00:11:02,867 It is generated from the measurement run, so it cannot drift from the evidence. 127 00:11:02,867 --> 00:11:08,363 Training runs are scheduled here and they are the only thing allowed to hold the GPU for a long time. 128 00:11:08,363 --> 00:11:14,738 Everything else in the system checks whether training holds the card before it asks for it, and backs off if it does. 129 00:11:14,738 --> 00:11:22,267 That single rule is why labelling, embedding, depth and three dimensional lifting can all run on one desktop without fighting each other. 130 00:11:22,267 --> 00:11:26,345 Lineage connects a deployed model back to every label it learned from. 131 00:11:26,345 --> 00:11:33,200 If a label is found to be wrong, this answers which models saw it, and therefore which measurements are now suspect. 132 00:11:33,200 --> 00:11:35,379 Events are the moments worth finding. 133 00:11:35,379 --> 00:11:40,041 Hard braking, a cut in, a near miss, a pedestrian crossing against the signal. 134 00:11:40,041 --> 00:11:43,794 They are detected from ego dynamics and object tracks together. 135 00:11:43,794 --> 00:11:45,913 An honest caveat from building this. 136 00:11:45,913 --> 00:11:54,933 The dynamics signal in this corpus is noisier than the events it is being asked to find, so the detector is tuned to over report and let a human cut. 137 00:11:54,933 --> 00:12:04,200 Events are searchable by description, so you can ask for every time a rider overtook on the left while the vehicle was braking, and get frames rather than a report. 138 00:12:04,200 --> 00:12:06,871 This is the part I would show a sceptic first. 139 00:12:06,871 --> 00:12:10,551 The system proposes its own work, but it cannot approve itself. 140 00:12:10,551 --> 00:12:18,267 Every autonomous action is a proposal with evidence, a named rule that fired, and a revert button that actually undoes the writes. 141 00:12:18,267 --> 00:12:22,241 A class is signed off by sequential testing rather than a fixed sample. 142 00:12:22,241 --> 00:12:26,896 A clean class proves itself in far fewer verdicts and a bad one fails in a handful. 143 00:12:26,896 --> 00:12:28,826 Accept is deliberately conjunctive. 144 00:12:28,826 --> 00:12:40,067 The sequential test and the confidence interval must both agree before anything is accepted, while a rejection is allowed to fire early, because rejecting early is the safe direction to be wrong in. 145 00:12:40,067 --> 00:12:45,473 The agent layer plans labelling work, criticises its own output and reconciles conflicts. 146 00:12:45,473 --> 00:12:50,633 It works in chunks, and every chunk is one recorded run with the exact rows it wrote. 147 00:12:50,633 --> 00:12:54,134 Reverting a run is a real database operation, not a flag. 148 00:12:54,134 --> 00:12:55,879 Governance sits above all of it. 149 00:12:55,879 --> 00:13:02,407 There is one killswitch that stops every autonomous writer, and every automated path checks it before doing anything. 150 00:13:02,407 --> 00:13:11,467 Approval thresholds are set per action here, so a team can let the system merge duplicate tracks on its own while still requiring a human for an ontology change. 151 00:13:11,467 --> 00:13:16,351 Every long running piece of work is a job with a state, a progress figure and a log. 152 00:13:16,351 --> 00:13:23,000 Jobs that die are reaped rather than left claiming to run, which sounds obvious and took a real bug to get right. 153 00:13:23,000 --> 00:13:27,556 One chronological record of everything that changed, by machine or by person. 154 00:13:27,556 --> 00:13:31,333 This is what you read when a number moved and nobody knows why. 155 00:13:31,333 --> 00:13:35,132 The reasoner explains a decision in words, with the evidence attached. 156 00:13:35,132 --> 00:13:39,812 Why this class was blocked, why this lot was rejected, why this model did not promote. 157 00:13:39,812 --> 00:13:47,466 The words are generated from the same fields the gate read, so an explanation cannot flatter a decision that the evidence does not support. 158 00:13:47,466 --> 00:13:50,847 The data engine is the loop that decides what to label next. 159 00:13:50,847 --> 00:13:52,451 More data is not better data. 160 00:13:52,451 --> 00:13:57,666 The question is always which thousand frames would move the number that is currently stuck. 161 00:13:57,666 --> 00:14:04,466 Frames are embedded as vectors, 42,088 of them, so you can search the corpus by resemblance rather than by keyword. 162 00:14:04,466 --> 00:14:08,999 Show me more scenes like this one is a query the system can answer directly. 163 00:14:08,999 --> 00:14:21,732 Text search runs over classes, attributes, sessions and scenario descriptions, and combines with the vector search, so a natural request narrows to a set of frames you can send straight to a labelling job. 164 00:14:21,732 --> 00:14:24,988 Discovery looks for what the ontology has no word for. 165 00:14:24,988 --> 00:14:30,455 Clusters of frames that no class explains well are surfaced as candidates for a new class. 166 00:14:30,455 --> 00:14:35,799 This is how an Indian road corpus grows a vocabulary that a western ontology never had. 167 00:14:35,799 --> 00:14:36,665 Sievyx filters. 168 00:14:36,665 --> 00:14:42,421 Near duplicate frames, frames that are almost all sky, frames the camera auto exposure ruined. 169 00:14:42,421 --> 00:14:51,332 Removing these before labelling is the cheapest quality improvement available, because a duplicate frame costs a full label and teaches nothing. 170 00:14:51,332 --> 00:14:53,716 The long tail is the whole problem in India. 171 00:14:53,716 --> 00:14:59,648 Cattle on the carriageway, three people on one motorcycle, an auto rickshaw reversing, a hand cart at night. 172 00:14:59,648 --> 00:15:07,465 These classes are rare in any corpus and over represented in the situations that matter, and this page ranks them by how starved each one is. 173 00:15:07,465 --> 00:15:08,616 This closes the loop. 174 00:15:08,616 --> 00:15:18,518 A blocked gate produces a recall deficit per class, the deficit builds its own review batch, the batch produces labels, the labels retrain the model and the gate runs again. 175 00:15:18,518 --> 00:15:24,332 That loop cleared a class that had been stuck at zero point one four recall for five training cycles. 176 00:15:24,332 --> 00:15:27,720 A campaign is a labelling goal with a budget and a deadline. 177 00:15:27,720 --> 00:15:36,332 The system allocates it against the marginal value of a label per rupee, using the measured time a verdict actually takes rather than an assumed rate. 178 00:15:36,332 --> 00:15:44,138 Interesting moments are exported as OpenSCENARIO files, so a situation found in real footage becomes a test case a simulator can replay. 179 00:15:44,138 --> 00:15:52,465 Where the esmini simulator is installed, each exported scenario is replayed at export time and the bundle records whether it was actually valid. 180 00:15:52,465 --> 00:15:54,024 Corpus level analytics. 181 00:15:54,024 --> 00:16:00,684 Class balance, geographic coverage, time of day, weather, how much of each session is labelled. 182 00:16:00,684 --> 00:16:05,078 One number here is uncomfortable and deliberately left visible. 183 00:16:05,078 --> 00:16:08,266 Of 598,179 objects, 788 carry a human verdict. 184 00:16:08,266 --> 00:16:17,265 The machine has run far ahead of the people, and pretending otherwise would make every quality number on this tour meaningless. 185 00:16:17,265 --> 00:16:20,304 A dataset is built from a query rather than a folder. 186 00:16:20,304 --> 00:16:26,673 All night time frames containing a rider without a helmet, excluding the gold set, excluding synthetic frames. 187 00:16:26,673 --> 00:16:31,465 The result is frozen with a hash and a datasheet that states exactly what went in. 188 00:16:31,465 --> 00:16:36,951 Synthetic frames exist, built by pasting real instances of starved classes into real backgrounds. 189 00:16:36,951 --> 00:16:40,208 They are quarantined structurally rather than by a filter. 190 00:16:40,208 --> 00:16:48,322 A synthetic label carries its own state and its own source, so every reader that measures anything excludes them without having to remember to. 191 00:16:48,322 --> 00:16:51,865 No synthetic pixel can reach a validation split or a gold set. 192 00:16:51,865 --> 00:16:57,619 Indian data protection law applies to road footage, because faces and number plates are personal data. 193 00:16:57,619 --> 00:17:07,132 Faces and plates are detected and redacted, and the redaction is verified by re-running detection on the redacted image rather than trusting that the blur was applied. 194 00:17:07,132 --> 00:17:14,208 Access to unredacted imagery is a role, every access is logged with a reason, and the log is not deletable from the product. 195 00:17:14,208 --> 00:17:17,372 Aggregated releases carry a differential privacy budget. 196 00:17:17,372 --> 00:17:28,532 Counts are noised, map cells with fewer than ten sessions are suppressed entirely, and when the budget for a scope is spent the endpoint refuses rather than quietly returning a weaker guarantee. 197 00:17:28,532 --> 00:17:31,222 The same spine runs a physical security pack. 198 00:17:31,222 --> 00:17:32,750 Cameras, zones, incidents. 199 00:17:32,750 --> 00:17:39,291 It shares ingest, annotation, review, measurement and governance, and shares none of the road specific code. 200 00:17:39,291 --> 00:17:44,732 An import contract is enforced in the build so that the core can never reach into a pack. 201 00:17:44,732 --> 00:17:48,442 Incidents in the security pack are assembled from tracks and zones. 202 00:17:48,442 --> 00:17:51,084 There is no transcription anywhere in this pack. 203 00:17:51,084 --> 00:17:57,999 No text is read off a frame and no text content is stored, which is a deliberate scope limit rather than a missing feature. 204 00:17:57,999 --> 00:18:01,119 Exports go out in the formats teams already use. 205 00:18:01,119 --> 00:18:05,302 COCO, YOLO, KITTI, nuScenes, OpenLABEL and the scenario formats. 206 00:18:05,302 --> 00:18:13,799 Every bundle carries a datasheet with the query that built it, the class counts, the synthetic share and the privacy parameters. 207 00:18:13,799 --> 00:18:16,218 Forgyx takes a trained model to the edge. 208 00:18:16,218 --> 00:18:20,451 Export, quantise to eight bit, compile for the target and benchmark it. 209 00:18:20,451 --> 00:18:27,466 A model that is accurate and too slow has not solved the problem, so latency is a gate condition and not a footnote. 210 00:18:27,466 --> 00:18:28,546 Deployment is staged. 211 00:18:28,546 --> 00:18:33,733 A candidate runs in shadow beside the current model on the vehicle before it takes over anything. 212 00:18:33,733 --> 00:18:43,133 The distillation loop here trains a small student from a large teacher and repeats until it fits the latency budget, or stops after four rounds and reports that it could not. 213 00:18:43,133 --> 00:18:44,692 Deployed models report back. 214 00:18:44,692 --> 00:18:48,849 Latency, thermal state, frames dropped and where the model was uncertain. 215 00:18:48,849 --> 00:18:55,200 Uncertain frames come home and re-enter the labelling queue, which is the outer loop of the whole data engine. 216 00:18:55,200 --> 00:18:56,899 Sanyx watches the hardware. 217 00:18:56,899 --> 00:19:01,867 Device temperature, storage, camera health and clock drift across the fleet. 218 00:19:01,867 --> 00:19:03,735 Predicted failures ahead of time. 219 00:19:03,735 --> 00:19:07,295 A camera whose focus is slowly drifting or a drive filling up. 220 00:19:07,295 --> 00:19:15,467 A vehicle that records unusable footage for a week is more expensive than one that stops, because nobody notices until the labelling starts. 221 00:19:15,467 --> 00:19:17,901 Operations shows the system's own health. 222 00:19:17,901 --> 00:19:21,369 Queue depths, GPU memory, worker state and the killswitch. 223 00:19:21,369 --> 00:19:25,872 The resource rules here are the reason this whole tour runs on one desktop. 224 00:19:25,872 --> 00:19:32,200 Every heavy path takes a slot, works in batches, checks free memory between them and yields to training. 225 00:19:32,200 --> 00:19:38,671 Several people work the same corpus at once, with assignment, locking and comment threads on individual objects. 226 00:19:38,671 --> 00:19:46,600 The interface is translated into Hindi, Kannada and Tamil, because the people doing the labelling are frequently not working in English. 227 00:19:46,600 --> 00:19:52,050 Each person sees their own throughput, their agreement with control samples and their history. 228 00:19:52,050 --> 00:19:55,800 Measured, not estimated, and visible to the person it describes. 229 00:19:55,800 --> 00:19:57,336 That is the full system. 230 00:19:57,336 --> 00:20:08,158 What is real today is the ingest, the auto labelling, the review surfaces, the tracking, the measurement with honest intervals, the governance and the export path. 231 00:20:08,158 --> 00:20:20,115 What is thin today is human verdicts, with 788 of 598,179 objects ruled on, and measured ego pose, where the 154 poses from this KITTI import are the first the corpus has ever had. 232 00:20:20,115 --> 00:20:28,065 Both of those are visible on the pages you just saw, because a data engine that hides its own gaps is not a data engine. 233 00:20:28,065 --> 00:20:28,933 It is a demo.