Vision on the edge
TensorRT pipelines and visual SLAM running in real time under strict power limits.
SVP MACHINE LEARNING · INFOCUSP
Standard models
fail. That's where
I start.
Computer vision researcher and ML leader. Ten years across peer-reviewed research and production systems — non-standard imagery, scarce labels, tight compute, auditable outputs.
9×
faster SLAM on edge
+20%
F1 in production
21
papers, 320+ citations
750+
stars on tf-cnnvis
What I work on
/ four recurring problemsAuditable agentic systems
LLM-as-a-judge layers so every output can be audited rather than trusted.
Multimodal retrieval
Fusing vector-indexed image and text embeddings with keyword search.
Scarce-label regimes
One-shot detection for tiny targets, imbalance and almost no labelled failures.
Selected work
View all 6 →9× faster inference
Real-time visual SLAM on edge hardware
A state-of-the-art SLAM pipeline taken from 0.26 to 2.4 FPS on a Jetson Orin NX — a ~9× speedup under strict power and compute limits.
- TensorRT
- ONNX
- Jetson Orin NX
- PyTorch
+20% F1 score
Defect detection where models plateau
A 20% F1 improvement on production inspection data defined by tiny targets, severe class imbalance, and almost no labelled failures.
- TensorFlow
- TF Model Garden
- OpenCV
Auditable by design
Agentic generation in a regulated domain
Health-protocol generation with an LLM-as-a-judge review layer, built so that every output can be audited rather than trusted.
- LangChain
- Agentic orchestration
- LLM-as-a-judge
Writing
- 2026.03 Drinking from the Firehose: Surviving the Velocity of Information
- 2020.09 Why fastai?
- 2020.07 Rest in Peace, Banerjee Sir.