Case study — Computer vision · Sports & media
EngageMax: predictive pose detection at production scale.
Turning live video into engagement intelligence: a real-time human pose estimation platform that runs continuously at broadcast frame rates — accurate enough to trust, efficient enough to scale.
Domain
Sports, media & engagement analytics
Engagement
Product engineering — end to end
Delivery
Chennai hub, global operation
Status
Live in production
At a glance
30 FPS
Real-time inference on live video
<100ms
End-to-end frame latency target
24×7
Continuous production operation
1 team
Model, pipeline & platform — one squad
The context
Engagement analytics needed eyes, not questionnaires.
Understanding how people physically respond — posture, motion, attention — has traditionally required manual review or intrusive instrumentation. Our client needed engagement measured continuously from ordinary video, at the speed of the event itself.
The answer was EngageMax: a computer-vision platform that detects and tracks human poses in live streams and converts movement into engagement metrics in real time. Shenll designed and built the full system — models, video pipeline, and analytics platform.
The challenge
Real-time or nothing
Pose estimation is compute-hungry; the value evaporates if analysis lags the live moment. Broadcast frame rates were the floor, not the goal.
Accuracy under messy reality
Crowds, occlusion, variable lighting, and camera angles — the model had to stay honest outside the lab.
Cost that scales sublinearly
GPU inference around the clock can bankrupt a product. Efficiency was a design requirement, not an optimization pass.
From detection to meaning
Raw keypoints aren’t insight. The platform had to translate skeletal motion into engagement metrics stakeholders act on.
What we built
A vision pipeline built like broadcast infrastructure.
01
Pose estimation engine
Deep-learning models detecting multi-person skeletal keypoints per frame, tuned and validated against real venue footage — with tracking that keeps identities stable through occlusion and crossings.
02
Real-time video pipeline
A streaming architecture that ingests live feeds, batches frames for GPU efficiency, and sustains 30 FPS end to end with sub-100ms latency budgets per stage.
03
Engagement analytics layer
Movement patterns aggregated into engagement scores, trends, and moment-level highlights — surfaced in live dashboards and post-event reports.
04
Production MLOps
Model versioning, drift monitoring, and evaluation harnesses that gate every model release; the same CI/CD discipline we apply to any enterprise system.
05
Predictive extension
Beyond detection: temporal models that anticipate movement and flag emerging moments before they peak — the “predictive” in predictive pose detection.
Architecture & stack
The results
Vision AI that earns its production badge.
EngageMax runs as continuous production infrastructure — not a demo reel. It anchors Shenll’s computer-vision practice and the engineering pattern we now apply across inspection, safety, and analytics use cases.
Broadcast-speed inference, sustained
Real-time pose analysis at 30 FPS with stable latency — during live events, not just benchmarks.
Reliability as a feature
Continuous operation with monitored model health, automated alerts, and rollback paths.
A reusable vision platform
The pipeline pattern now accelerates every Shenll computer-vision engagement, from factory QC to safety monitoring.
Insight, not just detection
Engagement metrics that stakeholders read at a glance — skeletons turned into decisions.
Explore the capabilities behind this work
Have a vision problem hiding in your video?
Detailed architecture reviews and extended write-ups are available under NDA. Tell us your challenge — we’ll show you how we’d approach it.