192 hours of dairy video labeled in a weekend

Nexa builds smart implants for cattle. Tiny, rugged devices that monitor animal health, behavior, and performance in real time. Their mission: help ranchers catch disease early, streamline breeding management, and maximize herd productivity while minimizing costs.

Behind the hardware sits a perception model trained on hours of in-barn cattle video. The quality of that model, and ultimately the device, depends on the quality of the labels feeding it.

"Generic labelers can tag a cow in a frame. They can't tell us if she's lame, in heat, off feed, or just lying down funny. We needed people who actually understand cattle."

Zarif Azher, CEO

The challenge: cattle video needs cattle-literate labelers

The behaviors Nexa cares about (early signs of mastitis, lameness, estrus, calving distress) look subtle on video. Distinguishing them takes someone who has spent real time around dairy cows.

The usual paths didn't work:

Generic labeling pools were too shallow. Off-the-shelf annotation services could draw bounding boxes, but couldn't reliably separate "sick" from "settling in for a nap." Quality fell apart on exactly the behaviors that mattered most.

Hiring and training in-house was too slow. Onboarding a labeler from scratch on dairy cattle behavior is a multi-week project. Nexa needed throughput on a model deadline measured in days.

Direct outreach to vets and farmers didn't scale. The right experts exist, but they don't live on labeling platforms; they're in clinics and barns.

"We had video. We had a labeling pipeline. What we didn't have was a quick way to put it in front of people who actually know what they're looking at."

Alvin Zhang, CTO

The solution: dairy vets and farmers, sourced and labeling in days

Working with Terac, Nexa stood up a vetted pool of dairy vets and active dairy farmers matched to their exact spec: credentialed, experienced with lactating herds, and able to start labeling that week.

Sourcing took days, not weeks. Onboarding piggybacked on expertise the labelers already had. They didn't need to be taught what a healthy cow looks like.

Over a single weekend, the pool labeled 192 hours of dairy video, tagging behaviors at the level of fidelity Nexa's model needed:

• Locomotion scoring and early lameness indicators
• Heat / estrus behavior
• Feeding and rumination patterns
• Calving and postpartum signs
• Social interaction and stress markers

"Terac dropped real dairy vets and farmers into our labeling tool in days. By Monday morning we had a labeled video set we couldn't have produced any other way."

Zarif Azher, CEO

The impact: cleaner training data, faster model iterations

Higher-fidelity labels. Subtle behaviors (the ones that drive the device's clinical value) were tagged by people who see them every day in real herds, not by labelers reading a rubric for the first time.

Faster iteration cycles. What had been a multi-week procurement and training problem collapsed into a sourcing turnaround measured in days. Nexa can spin up a new labeling push whenever the model needs more signal in a specific behavior class.

A repeatable expert workforce. The same pool can be re-engaged for re-labeling, edge-case review, and active-learning loops as the model evolves, without rebuilding sourcing each time.

The bottom line

For domain-heavy data, the bottleneck isn't tooling. It's getting the right people in front of the data quickly enough to matter.

By sourcing dairy vets and farmers in days, Nexa turned a weekend into 192 hours of high-quality labeled video, and a one-off labeling sprint into a repeatable engine for everything coming next.