← All projects

Medical / AgriTech · USA

Bovine Embryo Health Assessment — Multi-Model

An AI embryo-assessment platform for grade, stage, pregnancy, and sex prediction: a 30-second time-lapse clip goes in, and a production-ready health assessment comes out.

ROLEML Lead CLIENTEmgenisys · USA STATUSIn production

The challenge

Low expert agreement

Emgenisys research found that embryologists agreed on embryo stage only 59% on average; even highly experienced embryologists with 10+ years reached just 74.6% agreement.

Tiny, drifting subjects

Embryos are small and low-contrast; identity must survive the full-length video.

A production bar, not a demo

Every model had to clear deep evaluation analytics before approval.

The approach

01

Detection & tracking backbone

Driven to 99% across full-length videos — every prediction starts from a stable, correctly identified subject.

02

Spatial & temporal pattern recognition

The system analyzes morphology, motion, and change across the clip, combining frame-level visual cues with temporal patterns that are difficult to judge consistently by eye.

03

Goal-specific model design

Each model is designed around its primary goal — grade, stage, pregnancy, or sex — so the training objective, labels, and evaluation logic match the decision the model must support.

04

Calibration-aware evaluation & MLOps

Custom evaluation analytics, MLflow experiments, and CI/CD checks helped turn research models into approved production components.

Outcomes

99%

Detection & tracking

Embryo identity held across full-length videos.

4 models

Approved for production

Grade, stage, sex, and pregnancy prediction.

30 seconds

Clip to assessment

From raw video to a complete embryo readout.

Stack PyTorch Spatial-temporal ML Detection Tracking Fine-tuning MLflow Calibration losses CI/CD