A 38-engineer Sydney Pyrmont AV startup (Pirrama Road address, Series B raised early 2025, building ego-motion driving policy for a Tier-1 OEM in Melbourne) had 12,400 raw egocentric driving clips sitting in cold storage — front-camera + LiDAR + IMU + radar, 30 fps, 9-second windows, captured across Sydney CBD, the M5 East tunnel, and the Eastern Suburbs school run. The autonomy lead needed every clip labelled to the OEM’s ASIL-B traceability spec by the end of sprint 14, eleven calendar days out. The internal labelling pod (4 contractors in Brisbane) was running at 280 clips per day. Math says they ship 3,080 of 12,400 by the deadline. 75% of the dataset arrives at the OEM late, the homologation slot slips a quarter, the Series C diligence question becomes “why isn’t the training set ready?”.
This post is the worked example of what the AB7 Whitefield robotics-data pod actually ran, what it cost the Pyrmont startup ($0.62 per labelled clip, all-in), and what the ASIL-B evidence trail looks like when the dataset gets handed to the OEM.
What the deployment actually looks like
A 14-person AB7 Whitefield pod out of the Bengaluru office (Whitefield, ITPL Main Road, 4 km from EPIP Zone) — 11 annotators trained on multi-modal AV data, 2 QA reviewers ISO 26262-aware, 1 named ops lead (Ramya, CSCP-certified, four years on Waymo-spec data prep at a prior vendor). Tooling stack: Scale Nucleus for the dataset orchestration, CVAT 2.7 for the 2D + 3D cuboid annotation, Foxglove Studio for ROS bag inspection on the LiDAR + IMU + radar synchronisation, and a custom traceability ledger writing every annotator decision to a Git-backed audit log signed against the OEM’s ISO 26262 process ID.
Cost: $0.62 per labelled clip for the full multi-modal pass (front camera 2D bbox + LiDAR 3D cuboid + behaviour-tag triplet + ASIL-B trace ID). 12,400 clips × $0.62 = $7,688 fixed-scope, 11-day delivery. AU pricing is A$11,200. The internal Brisbane pod’s all-in burn rate worked out to A$1.72 per clip (4 contractors × A$120/day × 11 days = A$5,280; output 3,080 clips at most). Whitefield delivers 4× the dataset for 65% of the per-clip cost.
What’s in scope per clip:
- Front-camera 2D bounding boxes on 7 object classes — vehicle, pedestrian, cyclist, traffic light, road sign, lane marker, ground (drivable / non-drivable mask)
- LiDAR 3D cuboid annotation — same 7 classes, time-synced to the camera at 30 fps with sub-100ms drift tolerance verified per clip
- IMU + radar cross-check — annotator flags any clip where IMU and radar disagree on ego-vehicle yaw rate by more than 2°/sec (these flags go to a separate QA pool, not into training)
- Behaviour-tag triplet — the ego-vehicle’s intended manoeuvre (lane-keep / lane-change / merge / yield / brake-and-hold), the surrounding scene context (residential / arterial / motorway / school-zone / tunnel), and the rare-event marker (pedestrian-from-occlusion / cyclist-on-shoulder / emergency-vehicle / construction-zone — the OEM’s curated 14-tag rare-event taxonomy)
- ASIL-B trace ID — every annotation decision gets a SHA-256 hash of the (annotator-id, timestamp, decision-vector, tool-version) tuple, written to the Git audit log. The OEM can reconstruct who labelled what, when, with which tool version, for every clip in the dataset, for the regulator’s 10-year retention window
What’s out of scope and the buyer keeps in-house: the actual policy training (PyTorch on the buyer’s GCP), the OEM-side ISO 26262 hazard analysis (the buyer’s safety engineer signs that), and the homologation submission (the buyer’s regulatory consultant in Canberra files it).
The 11-day delivery window
Day 1-2 (Wed-Thu). Data-processing addendum + IP NDA signed. SSO into the buyer’s Scale Nucleus workspace provisioned to all 14 named pod members. Calibration round on 60 representative clips (20 CBD, 15 M5 tunnel, 15 Eastern Suburbs school run, 10 night-time wet-road). Inter-annotator agreement (IAA) on the 7 object classes: 94.2%. IAA on the 14-tag rare-event markers: 79.4% — below the 85% threshold. Two rare-event tags (pedestrian-from-occlusion, cyclist-on-shoulder) get a tighter definition pass on a Friday morning call with the buyer’s autonomy lead. Re-tested. Rare-event IAA climbs to 88.1%. Cleared to scale.
Day 3-9 (Fri-Thu). Production labelling at 1,520 clips/day net of QA rework. The 11 annotators on the camera + LiDAR + behaviour-tag pass. The 2 QA reviewers on 100% double-review of rare-event flags + 20% double-review of standard 7-class annotations. The Brisbane contractor pod runs in parallel on a separate 1,200-clip subset for the buyer’s Series C data room side-by-side.
Day 10-11 (Fri-Sat). Final QA pass on 62,000 annotation slots (12,400 clips × 5 attribute slots). 10% blind re-label sample shows 1.9% disagreement; 71% of disagreements are camera-bbox tightness within 4 pixels (informational per the OEM spec); 357 slots (0.58%) get corrected with chained audit log entries. Handoff: 87 GB dataset ZIP via buyer’s S3, ASIL-B trace ledger as a signed Git bundle. The buyer’s Pyrmont safety engineer runs git log --pretty=format:%H against the index — every hash chains cleanly. ISO 26262 audit-readiness confirmed in writing.
The Brisbane vs Whitefield side-by-side
The buyer ran a deliberate side-by-side: 1,200 clips to the Brisbane internal pod, the same 1,200 clips to AB7 Whitefield. The autonomy lead built a side-by-side slide for the Series C data room. The numbers:
| Metric | Brisbane internal | AB7 Whitefield | |—|—|—| | Throughput (clips/day) | 280 | 1,520 | | All-in cost per clip | A$1.72 | A$0.96 (US$0.62) | | Camera-bbox IAA | 91.8% | 94.2% | | LiDAR 3D cuboid IAA | 86.0% | 91.7% | | Rare-event tag IAA | 73.5% | 88.1% | | ASIL-B audit log present | No | Yes (SHA-256 chained) | | Delivery slip | 11.4 days late | On-time, day 11 17:00 IST |
The Brisbane pod isn’t bad. The Brisbane pod is one ops lead short of being audit-ready — there’s no ISO 26262-aware QA layer, and the rare-event taxonomy needed a tighter definition pass that nobody had time to run. AB7’s value here is not “cheaper offshore labellers”. The value is the named ops lead who has shipped this exact format before, and the ASIL-B trace ledger that the OEM’s safety engineer can audit.
What this is not
This is not “anyone with a CVAT licence can do this.” Egocentric multi-modal AV data needs an ops lead who has seen Waymo-spec, Cruise-spec, or Wayve-spec data formats before. There are roughly 40 of those people in India. AB7 employs 3 of them.
This is not “label everything, ship to the OEM, hope for the best.” The 14-tag rare-event taxonomy got re-defined on day 2 because IAA was below threshold. Most vendors would have shipped on day 11 with the bad taxonomy and let the OEM find out at sprint 16 when the training run produced a brittle policy on cyclist-occlusion scenarios.
This is not for buyers without an OEM-side safety engineer. The ASIL-B trace ledger only matters if there’s someone on the buyer’s side who knows what to do with it. AB7 will say so on the scoping call.
What the autonomy lead in Pyrmont said in the post-mortem
“We saved roughly A$13,000 against the Brisbane pod’s run-rate plus the consultant we would have paid to clean up the rare-event taxonomy in week 3. The ASIL-B trace ledger is the thing the OEM’s safety engineer in Melbourne actually opened first. None of the other three vendors we evaluated had that as a standard deliverable. The Whitefield ops lead joined our standup on day 2 at 09:30 Sydney = 04:00 IST — she didn’t need to, but it set the tone for the rest of the sprint.”
The objection a Sydney AV startup CTO usually raises
“Whitefield is 4.5 hours behind Sydney. Won’t our daily standup land at 04:00 in Bengaluru and break the team?” The honest answer: only the ops lead joins the daily standup. The standup at 09:30 Sydney = 04:00 IST is on Ramya’s calendar; the rest of the pod works the 10:00-18:30 IST window, which overlaps with 14:30-23:00 Sydney for the autonomy lead’s afternoon block. Afternoon block = synchronous rare-event-tag clarifications. Morning block (Sydney) is async — Ramya leaves a tagged Slack thread before midnight Bengaluru, the Sydney lead reads it at 08:00 with coffee. QA double-review runs overnight Sydney time, so the autonomy lead walks in at 08:30 to clips already QA’d, not still in flight.
What happens in the first 60-minute call
Ashok Benial (founder of AB7, Calendly link below) takes the scoping call. Three things on a 60-minute call: (1) the actual scope and timeline — clip count, modality mix, OEM traceability spec, deadline; AB7 returns the per-clip price and throughput math. (2) The taxonomy preview — Ramya joins (14:30 Sydney = 09:00 IST), walks a 10-clip pilot of the buyer’s existing taxonomy, names the 2-3 definitions she’d tighten before scaling, shows the ASIL-B trace ledger format. (3) The deployment plan — named pod members, ISO 26262 training certs, IP NDA terms, cutover Monday. The CTO leaves with a written plan, three LinkedIn URLs, and a Wednesday start date.
Book the 60-minute scoping call →
Related reading
- AB7’s AI & Robotics Services hub — 10 sub-services including robotics data ops, HITL review, RAG curation, agent QA: /ai-services
- AB7’s automotive & mobility industry page — engineering CAD, vehicle data ops, EV/mobility software, AV training-data: /services/by-industry/automotive-mobility
- The robotics data + ops case study (anonymized AV training-data deployment): /case-studies/robotics-data-operations-av
- AB7’s full pricing page — FTE rates, multi-discipline team rates, per-task pricing for labelling pods: /pricing
SCHEMA INSTRUCTIONS (if WP’s RankMath / Yoast doesn’t auto-emit BlogPosting JSON-LD)
Paste this into the WP post’s “Custom Fields” or RankMath schema block. The VideoObject block is included as an optional second @graph entry — only emit it if the OG endpoint or WP post includes the 28-second labelling-tool screen-capture video. If no video ships, drop the VideoObject entry and keep just BlogPosting.
Validate at https://search.google.com/test/rich-results before publishing. Expected outcome: 1 BlogPosting object detected + 1 VideoObject object detected, 0 errors, warnings on image.width/image.height and VideoObject.contentUrl OK (those resolve once the OG endpoint returns a real PNG with EXIF and the MP4 file lands at the contentUrl). If the labelling-demo video is not shipping with this post, delete the VideoObject from the @graph and re-validate — expected outcome then: 1 BlogPosting only, 0 errors, same image warnings.