Academic & Professional Experience

Leading development of a full detection, tracking, and risk-alert system for human-wildlife conflict, mentored by Prof. Sukanta Saha.

• Trained and evaluated a PyTorch-based YOLO11 detection model and ByteTrack multi-object tracker on a self-annotated wildlife dataset, reaching 91.2% mAP@50 and 93.1% precision.
• Owned dataset curation, training, quantitative evaluation (mAP, precision/recall, ablation studies), and edge deployment end to end.
• Designed a spatial risk-scoring model that translates raw detection output into an actionable, interpretable alert for a real user.

Designed AI-driven decision support systems and scheduling optimization models for intelligent infrastructure.

• Architected a 5-Layer Universal Scheduling Engine modeling complex multi-variable operational constraints.
• Translated compliance rules (workload balance, fatigue limits) into integer programming variables using Google OR-Tools (CP-SAT solver).
• Incorporated data pruning heuristics using custom spatial distance matrices to scale performance, yielding allocation matrices in <2.3s.

Built core pipelines for a UAV video intelligence and geospatial detection platform.

• Built a dual-pipeline telemetry system decoding embedded MISB KLV metadata from drone video and, for feeds without it, extracting flight data via OCR on the HUD overlay, converging both into one schema.
• Implemented pixel-to-ground target geolocation using ray-casting against DEM/terrain data, plus a 3D flight viewer for visualizing tracked targets over terrain.
• Trained a U-Net (ResNet-34) segmentation model on satellite imagery for infrastructure detection, reaching 92.0% validation IoU after 60 epochs, and integrated audio-based target detection using CLAP embeddings alongside vision-based detection.

Researched hardware-enforced secure computing backend systems using Intel SGX enclaves.

• Programmed low-level ECALL and OCALL boundaries to minimize the Trusted Computing Base (TCB) for private document pipelines.
• Processed workloads up to 2.86 GB of encrypted data inside a hardware-isolated Intel SGX enclave with only ~16 MB of trusted memory (EPC), using memory paging/swapping to fit far more data than the enclave's native capacity without breaking isolation.
• Deployed and benchmarked enclave-bound cryptographic keys (AWS KMS) on Azure Confidential Computing VMs (DCsv2/DCsv3).

Created distributed wildlife analytics systems for resource-constrained edge environments.

• Curated a 4,000-image polygon segmentation dataset utilizing Grounding DINO and SAM (Segment Anything Model) as a teacher pipeline.
• Applied knowledge distillation to train a student YOLO-seg model, quantizing to INT8 for a 75% size reduction.
• Decomposed processing workloads for edge deployment on Raspberry Pi 5 running at 5-8 FPS on CPU and ESP32-CAM nodes.