Seeking collaboration/advice: Real-time digital twin of workplace stress & cognitive productivity using behavioral biometrics (WhiteMatter platform)

AI & Data Science NVIDIA AI Workbench
1

Please describe your issue or request: (tick the boxes after creating this topic):

Hi everyone,

We’re Erasys Ltd (UK-based NVIDIA Inception member, recently accepted), building WhiteMatter: an edge-first AI platform that creates a real-time “digital twin” of an employee’s cognitive state (stress, cognitive load, focus depth, fatigue, emotional valence) using only passive behavioral biometrics from standard laptops—no wearables, no cameras, no raw keystrokes logged.

Current stack & performance

  • Lightweight agent on Windows/macOS/ChromeOS

  • On-device multimodal models (currently PyTorch → ONNX → TensorRT) running at ~18–25 ms latency on RTX 4000/5000 Ada laptops

  • 89% accuracy against clinical ground truth (500k+ labelled hours)

  • All sensitive processing stays on-device; only anonymized cognitive scores go to the cloud

What we want to achieve next (and where we’d love the community’s help)

  1. Push inference latency below 10 ms on RTX enterprise GPUs using latest TensorRT 9.x / TensorRT-LLM tricks and 8-bit / 4-bit quantization without accuracy drop.

  2. Explore NVIDIA’s BioNeMo / NeMo Curator frameworks to pre-train a foundational “human-digital interaction” model on >5M hours of upcoming NHS/university pilot data.

  3. Build a real-time “digital twin” visualization dashboard in Omniverse / Isaac Sim that shows live population-level cognitive heatmaps for entire office floors (privacy-preserving of course).

  4. Federated learning across enterprise fleets while keeping everything on-device (any success stories with NVIDIA FLARE in corporate environments?).

Happy to share anonymized model architectures, sample inference pipelines, and early TensorRT benchmarks with anyone interested in collaborating or giving optimization advice.

If you’ve pushed similar low-latency behavioral models to the edge with TensorRT, worked with BioNeMo on non-genomic time-series data, or built Omniverse-based workforce analytics twins, I’d love to connect!

Thanks!
Ibrahim @ Erasys (NVIDIA Inception)
https://erasys.co.uk| ibrahim.mukherjee@erasys.co.uk

Please tick the appropriate box to help us categorize your post
Other