Projects

Edge-AI Driver-Monitoring Platform (AutoAware AI) (2024 - 2025)

Conceived the company, built a functional MVP prototype, and secured non-dilutive funding.
Developed quantisation-aware vision models that recognise 17 driver-safety behaviour classes from an 800,000-image dataset (e.g., phone use, drowsiness, distraction-related actions).
Delivered real-time inference on low-cost edge-AI boards—balancing speed, accuracy, and hardware cost—via INT8 quantisation and input-resolution optimisation, with negligible latency and very low false-alarm rates.
Implemented an on-device driver-assistance prototype that provides real-time alerts; pilot testing across multiple vehicles demonstrated high detection precision. [Video1] [Video2] [Video3]

New Perspectives for the Metaverse-as-a-Service (2023)

Wrote a comprehensive white paper on privacy, security and human-centric design for MaaS; compared wireless-access models, learning pipelines and data-governance schemes.
Proposed an edge-computing and blockchain architecture that telecom partners now use as an implementation roadmap.

Deep Learning with PyTorch (2023)

Designed and taught a full-semester course for EE graduate students: classification, custom datasets, experiment tracking, model deployment, transfer learning, ViT, GAT, GANs, RL.
Released all Colab notebooks and code on GitHub for open access.

Udacity Self-driving Car Engineer Nanodegree Program (2018)

Built CV/DL pipelines (OpenCV + TensorFlow/Keras) for lane detection, traffic-sign recognition and object classification; > 95 % accuracy.
Implemented LiDAR+radar EKF sensor fusion for real-time vehicle tracking.

Traffic Congestion Detection From Camera Images (2017)

Trained YOLO and DCNN models (91.5 % / 90.2 % accuracy); benchmarked classical ML (SVM f₁ = 86.7 %).

Computer-Vision Challenges

Lyft Level-5 Perception (2019): Ranked 14ᵗʰ/300+ with MobileNet-based semantic-segmentation model (CARLA & KITTI datasets).
NVIDIA AI City (2019): Achieved 96 % vehicle-detection accuracy; estimated speeds via refined vanishing-point geometry.

Crowdsourced & Probe-Data Analytics (2018)

Waze Evaluation: Clustered irregular incident reports using Sparse PCA; compared Waze, INRIX and Wavetronix feeds.
1 TB Radar/Probe Performance Monitor: MapReduce + Apache PIG pipeline; built Python spatiotemporal anomaly networks and Tableau dashboards.

Game-Day Mobility Studies (2019)

Devised Extended-EigenSpot hotspot algorithm; used DBNs to forecast hotspot onset/locations.
Quantified speed deviations and probe-data coverage for collegiate football events.

Real-Time Traffic Incident Detection (2017)

Adapted Dynamic Time Warping to compute optimal warping paths; achieved real-time incident flagging on streaming data.