Srivatsan Sarvesan

A benchmark of forward-prefill and single-token KV-cache attention across six NVIDIA GPUs, asking whether the fastest backend on one device is still the fastest on another. A winner counts only when it clears both a bootstrap confidence interval and a 10% margin; under that bar the backend changed in 27 of 64 separated prefill comparisons across the Ampere, Ada and Hopper parts.

A wearable assistive system pairing a Raspberry Pi belt with an Android app that runs Gemma-3n vision-language inference, ArcFace face recognition, and sensor-fusion fall detection on-device, behind a fully voice-driven interface.

A modular agentic AI for chest X-ray interpretation built on GPT-4o-mini, integrating segmentation, classification, grounding, VQA, and report generation with a RAG-backed vector DB for retrieval-augmented, human-in-the-loop reasoning.

Algorithms for reconstructing 3D coronary geometry from X-ray angiograms, improving stenosis-detection accuracy and cutting computation time via pseudo-transient vFFR.

A multimodal multitask framework for Alzheimer's classification (NC/MCI/AD) via joint MRI–PET fusion with auxiliary modality reconstruction.

Leading the churn and patient-retention analytics workstream for remote hypertension and diabetes programs, analyzing longitudinal patient histories to derive data-driven at-risk and behavioral-churn thresholds, disengagement taxonomies, and early warning triggers. Translating these patterns into clinical and operational intervention protocols for identifying and re-engaging patients before churn. In parallel, I built a cuffless BP estimation pipeline from raw PPG, using autonomous AutoML agents to optimize few-shot subject-specific calibration from as few as ~10 samples.

Building a multi-view computer vision system for cricket analytics using synchronized high-frame-rate iPhone video to reconstruct ball trajectories and estimate how shots played in practice nets would carry and land under different ground settings. My work spans CVAT-based dataset curation, YOLO/CoreML training and benchmarking, single-view ball detection, two-view synchronization and tracking, camera calibration, and 3D trajectory reconstruction.

Developing efficient search and memory architectures for autonomous data-engineering agents, enabling knowledge reuse across tasks while reducing time and inference/API cost. My research explores cross-task memory, multi-agent coordination, self-evolving workflows, and test-time scaling for data-centric autonomous AI.

Developed algorithms for 2D-to-3D coronary artery reconstruction from X-ray angiography, improving stenosis-detection accuracy and reducing computation time via pseudo-transient vFFR. Currently building deep learning models for laparoscopic cholecystectomy: video-based CVS classification and hepatocystic anatomy segmentation.

Built a DenseNet-based multi-scale pyramidal fusion network for retinal vasculature and FAZ segmentation in OCTA images to aid diabetic-retinopathy diagnosis. Developed a multimodal, multitask framework for Alzheimer's classification (NC, MCI, AD) integrating MRI and PET data.

Designed a flexible hexagonal patch antenna on PDMS for sub-6 GHz 5G wearables, achieving 3.2 dB peak gain with a validated low-cost fabrication process. Compared Random Forest, XGBoost, CatBoost, and LightGBM to optimize antenna parameters from full-wave EM simulation data.

Developed a spatial-attention segmentation model to detect spatter-particle defects in X-ray Computed Tomography, improving defect-detection accuracy for additive-manufacturing quality control of metal parts.

Collected and analyzed single-channel sEMG during hand-reaching tasks to extract muscle-activation patterns using signal processing and ICA. Modeled sEMG signals to study motor-unit recruitment dynamics during voluntary contraction.

GPA 9.18/10 · Minor in CSE