I build deployable computer vision and edge-to-cloud AI systems for real-world sensing, emphasizing robustness, scalability, and reliability. My work bridges model development and field deployment—designing end-to-end pipelines from data capture and processing to inference, monitoring, and human-in-the-loop validation.


Selected projects

MosquitoAI dashboards

MosquitoAI — AI Dashboards for Mosquito Biology & Species Inference

Jan 2021 – Present · Supervised by Dr. Sriram Chellappan · University of South Florida

End-to-end computer vision pipeline and interactive dashboards for mosquito surveillance, supporting species and biological inference from heterogeneous imagery.

  • Developed multi-task deep learning workflows across adult and larval stages for classification and biological inference.
  • Established reproducible training, validation, and benchmarking procedures to support deployment-oriented performance.
  • Integrated explainability outputs to support human verification and operational trust.
  • Packaged results into dashboards designed for field partners and applied surveillance workflows.
Edge-to-cloud smart mosquito trap

Edge-to-Cloud Smart Mosquito Trap for Automated Vector Monitoring

Apr 2023 – Present · Supervised by Dr. Sriram Chellappan · University of South Florida

Field-deployable sensing system that captures multi-focus imagery on-device and synchronizes data to the cloud for automated detection, analytics, and alerting.

  • Designed an end-to-end pipeline from edge capture (multi-focus) to cloud storage, metadata logging, and model inference.
  • Built automated workflows for extracting single-mosquito crops via localization and preparing training-ready datasets.
  • Enabled scalable monitoring through standardized capture formats, metadata schemas, and deployment-ready processing.
  • Implemented robust evaluation to quantify performance under real-world conditions (clutter, occlusion, domain shift).
Valvular disease classification from heartsound data

Valvular Disease Classification from Heartsound Data

Feb 2020 – Jul 2020 · Supervised by Dr. Taufiq Hasan · Bangladesh University of Engineering and Technology

Computer-aided analysis of heartsound recordings for reliable cardiac abnormality detection under noise and sensor variability.

  • Developed preprocessing and feature pipelines for heartsound signals under real-world noise and acquisition variation.
  • Trained and benchmarked deep classifiers for abnormality detection across multiple datasets.
  • Focused on reliability in challenging conditions (additive noise, device variability, domain mismatch).
  • Reported strong screening-aligned performance with improvements in discrimination metrics.
Dynamic ROI hand vein authentication

Robust Human Authentication via Dynamic ROI from Hand Vein Images

May 2019 – Dec 2019 · Supervised by Dr. Mohammed Imam-ul Hassan Bhuiyan · Bangladesh University of Engineering and Technology

Vision-based biometric authentication using dynamic ROI extraction from dorsal/palm hand vein imagery with robustness to pose and acquisition variability.

  • Implemented dynamic ROI extraction to reduce sensitivity to alignment, pose, and background variation.
  • Trained deep models for identity verification and evaluated robustness across acquisition conditions.
  • Built a streamlined pipeline suitable for integration into practical authentication workflows.
  • Emphasized reliability and security considerations relevant to real deployment environments.