CV — Explore by Skill

Edge 3D perception for warehouse detection and depth-aware navigation on Jetson-class hardware. Improved detection mAP 0.15 → 0.62 (synthetic; 0.58 real), cut end-to-end latency to 26 ms at 640×640, and combined depth + detection at 10 FPS on Jetson Orin NX.

Active Inference routing agent on the perception stack; JAX planning optimized from 7 min → 21 s. PyMDP saccading agent on an IoT camera and Isaac Omniverse simulation for warehouse routing evaluation.

Built Docai.tools: OCR → private RAG architecture (chunking, embeddings, on-prem deployment). SOTA VLM research for Drizz.dev and deep-learning strategy for Eventgraphia (image culling).

End-to-end ADAS perception with MOT on AWS; extended 2D detection to 3D bounding boxes. Production monitoring and alerting with Weights & Biases across model versions.

Docker/CI/CD MLOps for multi-TB aerial segmentation; Sacred/Omniboard across 20+ iterations (~40% faster experimentation). GPU vs TPU benchmarking to inform production infrastructure strategy.

Robustness/privacy research on CNNs, Lipschitz bounds, and DP threats, with containerized reproducible evaluation pipelines. Preprint: gradient masking and underestimated robustness threats of DP in deep learning.

DDPM synthetic adversarial datasets for CIFAR-10 robustness retraining (+8–10% ResNet18 accuracy).

“Towards smart and adaptive agents for active sensing on edge devices” — peer-reviewed publication on edge active sensing.

Urban pedestrian segmentation and LiDAR PoC, with an MLOps pipeline (Docker, experiment tracking, Go-task automation).

U-Net multi-coil MRI 3D reconstruction (3rd place globally); NeRF exploration for CT-from-X-ray. Published in Frontiers in Neuroscience.