Real-time guidance for human data collection
DriverSeat is a voice copilot that works alongside the people collecting your training data. It knows the task, it knows what you already have, and it turns what they say into clean tags while they work.
Multi-language. On-device or on-prem.
"Dropped the box on the second attempt, the flap was open and the gripper slipped on the tape."
DriverSeat
"Same SKU as the previous five? We have 40 of those."
Session
“There's a broken sign lying across the road, it's blocking most of the lane — I had to stop and wait for the car in the other lane to pass.”
DriverSeat: “Was it daylight and dry? I'll log the conditions. This is the first obstruction we have from this district.”
“Dropped the box on the second attempt, the flap was open and the gripper slipped on the tape.”
DriverSeat: “Same SKU as the previous five? We have 40 of those. If so I'll mark it low priority.”
“Longitudinal crack on the third insulator from the top, about ten centimeters, and the metal under it is rusted.”
DriverSeat: “Logged with location and time. Want a photo on this one before we move on?”
Every tag has a timestamp and follows your schema. It's text, so your pipeline can search it for pennies instead of re-watching hours of video.
Use cases
Operators on eight-hour shifts, judged on usable hours. Catch the bad episode at minute two instead of at review.
Drivers hunting edge cases across cities and countries. Structured events the moment they happen, instead of daily reports and button presses.
Inspectors describing defects hands-free. Speech becomes a record with location, type and severity.
Collectors who get rejected without knowing why. Tell them while they can still fix it.
The problem
Drivers, teleoperators and field collectors get the task through two or three people. They don't know what's already covered. They tag with a button or a voice memo. Quality control happens after the shift, when the hour is already paid for and the footage is already stored.
Every company we studied runs QA the same way. Record, upload, review, reject. Nobody guides the collector while the data is being made.
Better instructions cut severe labeling errors by a third. More QA barely moves them.
0.03%
Share of daily driving that is a long-tail scenario.
Source: Waymo, WOD-E2E
5M → 2M km
BMW collected 5M km of real driving. 2M km were relevant.
Source: BMW Group D3
4 of 8 hours
Minimum usable footage per shift Tesla demands from Optimus data collectors.
Source: Business Insider
<2%
Quality gain from vendors' internal QA.
Source: Rädsch et al., ECCV 2024 / Nature MI 2023
How it works
01
02
03
Session report
Morning shift · Site B
Coverage
Tags
Deployment
Speech and extraction run on the phone, headset, or edge device. Audio never leaves the collector's hands. Works offline and syncs later.
Models on your GPUs, inside your network. No third-party API in the loop.
For defense, regulated and export-controlled programs. Nothing phones home.
DriverSeat runs on open-weight speech and language models that we fine-tune and orchestrate. No vendor lock-in. When a better open model ships, you get it. Your task content, schemas and recordings stay yours.
Pilot
Weeks 1–2
We take your task briefs and tag schema and configure DriverSeat. You pick five to ten collectors.
Weeks 3–6
Collectors run real shifts with DriverSeat. We measure usable-data rate, tag completeness and time-to-tag against your current baseline.
Weeks 7–8
On-prem deployment on your infrastructure. You decide with numbers.
You define success before we start.
FAQ
Your collectors' real task, guided end to end.