Subhankar Sen

Leading an R&D team of AI/ML engineers in architecting and deploying CRAVE (Cardiac Rhythm AI Visualization Engine), a production-grade, device-agnostic ECG arrhythmia detection platform processing 432,000+ beats from multi-day Holter recordings and achieving 92%+ real-world clinical accuracy on unseen patient datasets. Engineering a 1D pre-activation ResNet rhythm model (16 residual blocks, global temporal pooling), delivering 97% accuracy in Normal vs. classification on 10-second ECG segments. Designing and implementing BeatNet, a hybrid deep learning architecture combining multi-scale convolutions, Gaussian region pooling, and lightweight MLP layers, achieving 94.5% accuracy across N/S/V/F beat classes with adaptive rhythm orchestration logic. Orchestrating a comprehensive 6-stage ML pipeline integrating ConvBiLSTM-based R-peak detection, beat classification, PQST segmentation, arrhythmia sub-classification, and morphology-based clustering achieving >85% concordance with expert clinical interpretations. Developing an advanced ECG denoising framework leveraging NeuroKit2, PyWavelets, EMD, SSA, morphology-driven Wiener filtering, dynamic Kalman filtering, and DTW-based alignment, enhancing P-QRS-T wave delineation and diagnostic accuracy. Building a device-agnostic preprocessing framework using MNE for EDF parsing across multiple manufacturers, with automated resampling, sliding-window analysis, gap detection, and daily batch processing for 8,600+ ECG segments per day. Resolving critical memory and performance constraints by reducing RAM consumption from 15 GB to <3 GB and cutting processing time for 2-day recordings to under 20 minutes through adaptive caching, cleanup strategies, and dynamic batch scaling. Optimizing cloud deployment workflows by reducing Docker image size from 4 GB to 1.8 GB, applying multi-stage builds, layer caching, and PyTorch cleanup, while integrating real-time performance monitoring for GPU, VRAM, CPU, and latency. Architecting a fault-tolerant AWS infrastructure using Step Functions, AWS Batch, EKS, ECR, S3, and CloudWatch, ensuring parallel containerized execution, automatic recovery, and continuous observability in large-scale clinical data processing.

Architected and deployed both static and dynamic kitting inference pipelines, automating kit processing and scan validation workflows, and delivering high-precision validation reports for internal and client systems achieving a 98% kit PASS rate and enhancing operational accuracy for a leading US-based machine vision filters manufacturer. Developed and implemented an end-to-end YOLOv8 segmentation + Google OCR service bundle for bolt serial number detection, monitoring the full lifecycle from Proof of Concept (POC) to Phase-1 production rollout, ensuring reliable, real-time performance in industrial environments. Applied unsupervised clustering algorithms, including Gaussian Mixture Models (GMMs) and K-Means, to detect and classify color anomalies in vitrified floor tiles, delivering a tunable anomaly detection system that enabled customizable precision in differentiating between original and faded variants. Integrated and enhanced OCR pipelines by combining EasyOCR for scene text recognition and ZXing for barcode/QR parsing, while replacing the default CRAFT text detector with a YOLOv8-based model trained from scratch using synthetic cross-domain datasets, resulting in significant accuracy improvements for industrial scene text detection. Designed and executed secure data exchange protocols between ML inference and application servers using PPP tunnels, ensuring end-to-end encryption, low-latency communication, and full compliance with enterprise data privacy standards. Developed a frame reader module utilizing Baslers pylon SDK, enabling dynamic configuration of camera parameters through Feature Stream file loading, improving usability and alignment with industrial imaging and control system workflows. Evaluated and fine-tuned YOLOv8 models across multiple industrial AI applications including kitting, PPE compliance, hazard detection, and worker safety, enhancing model precision and improving operational throughput through continuous retraining and optimization. Built and deployed an advanced face anonymization tool compliant with European data protection regulations (GDPR), ensuring ethical AI implementation and safeguarding worker identities. Led end-to-end testing and validation of ML inference servers, app servers, and data manager Docker releases; executed edge case simulations, authored unit tests, and ensured seamless deployment across x86 cloud and ARM edge environments. Mentored and guided three final-year engineering interns through applied AI projects, fostering technical growth, collaboration, and effective knowledge transfer within the ML engineering function.

Standardized histopathology datasets by converting them into the COCO JSON format, enhancing interoperability for machine learning applications. Developed innovative training methodologies by creating mosaic cut-and-mix AOIs with updated nuclei annotations to diversify training datasets. Adapted and optimized the StarDist model for multi-task learning, enabling simultaneous instance segmentation and classification of multi-organ nuclei types. Conducted detailed spatial analytics through cell statistics analysis, utilizing advanced data formats such as GeoJSON and HDF5 for comprehensive insights. Achieved a notable reduction in inference times for advanced models, contributing to more efficient processing and analysis of complex datasets.

Designed a multi-task learning model aimed at improving surgical scene understanding, focusing on tool-tissue interaction detection and instrument segmentation. Collaborated with cross-disciplinary teams to integrate advanced machine learning techniques into surgical robotics applications.

Developed a comprehensive deep learning architecture for Temporal Action Localization (TAL) in untrimmed videos, focusing on improving detection accuracy. Reformulated existing action segmentation schemes to effectively address the TAL task without relying on complex proposal generation methods. Achieved a 40.3 mAP on the THUMOS14 dataset, surpassing the current state-of-the-art model by 2%, showcasing the effectiveness of the developed architecture.

Developed a computer-aided diagnosis (CAD) system utilizing feature extraction and enhancement techniques for breast lesion microscopy images. Contributed to the development of a CAD system that improved diagnostic accuracy for breast lesions, showcasing the impact of AI in healthcare.