GE HealthCare
AI Researcher — Advanced Technology Group
Voxel-Adaptive Denoising for High b-value Diffusion MRI
- Invented gradient-informed structural prior that dynamically adjusts kernel geometry per-voxel, leveraging local tissue anisotropy to preserve sharp tumor boundaries whilst aggressively suppressing noise in low-anisotropy regions.
- Developed gradient-to-kernel mapping algorithm using k-means clustering of voxel-wise gradient magnitudes with monotonic assignment of discrete odd-integer kernel sizes — generating a per-voxel kernel map that autonomously adapts MPPCA denoising to local anatomical complexity.
- Delivered 18% SNR gain with 96% tissue boundary preservation vs. fixed-size MPPCA baselines (NYU-MPPCA, DIPY-MPPCA) on HCP data at b-values of 5000–10000 s/mm².
- Filed as Invention Disclosure at GE HealthCare. Accepted at EMBC 2026, Toronto.
MRI Acquisition Parameter Optimization in Fast Spin Echo (FSE)
- Replaced manual FSE protocol tuning with a physics-informed optimization framework for 2D and 3D sequences, automating echo train length, T2-weighting, and eddy current trade-off selection.
- Reduced scan setup time from hours to minutes across 4 clinical sites.
- Investigated critical FSE physics: echo train length optimization, refocusing pulse efficiency, eddy current compensation, chemical shift artifacts, and J-coupling effects in multi-echo sequences.
CTI-Based PSF Measurement & Distortion Correction in MRI
- Developed a Constant Time Imaging (CTI) framework for voxel-wise Point Spread Function (PSF) measurement — by holding excitation-to-readout time constant, T2* contrast is eliminated and the PSF is isolated, enabling direct quantification of distortions that conventional sequences conflate with signal.
- Designed a voxel-wise phase-encoding scheme that independently samples the local PSF at each spatial location, producing spatially varying distortion maps that capture tissue-dependent T2 relaxation differences, susceptibility gradients near air-tissue interfaces, and inter-voxel motion-induced phase corruption.
- Decomposed four major PSF-broadening mechanisms: (i) T2 decay–driven exponential apodization of k-space along the readout axis, (ii) susceptibility-induced B₀ inhomogeneity causing geometric warp and signal pile-up, (iii) chemical shift displacement between fat and water resonances in frequency-encode and phase-encode directions, and (iv) through-plane rigid-body motion introducing inter-shot phase inconsistencies.
- Built a correction pipeline that inverts the measured voxel-wise PSF prior to reconstruction, recovering spatial resolution in distortion-prone regions (orbitofrontal cortex, posterior fossa, air-tissue boundaries) where standard EPI protocols suffer significant blurring and geometric error.
Dell Technologies
Software Engineer Winter Intern — Dell Digital
ML-Driven Order Fulfillment Forecasting
- Designed configuration-aware scoring model and feature engineering pipeline processing 100K+ daily transactions via SQL Server.
- Identified low-probability SKU configurations that reduced idle inventory by 15% in A/B testing, extracting actionable business intelligence for supply chain optimization.
- Built automated drift detection triggering targeted data cleaning, reducing manual effort by 40% and enabling daily retraining for sustained model accuracy.
Dell Technologies
Software Engineer Summer Intern — Dell Digital
RAG-Based Enterprise Knowledge System
- Architected RAG pipeline with GPT4ALL, LangChain, and Pinecone vector store, engineering chunking strategy and embedding indexing over 10K+ internal documents.
- Achieved 92% retrieval accuracy (MRR@5) on technical documentation queries — 3× faster than legacy keyword search.
- Shipped production deployment with RBAC, audit logging, and latency monitoring. Adopted by 3 engineering teams for daily technical Q&A workflows.
ZKTeco Inc.
Research Intern — Biometric Systems R&D
Multimodal Biometric Authentication System
- Engineered multimodal biometric pipeline fusing finger vein and fingerprint data via Siamese Networks trained with contrastive loss, significantly improving spoof resistance through feature-level fusion.
- Designed preprocessing pipeline with CLAHE and ROI extraction to enhance vein pattern visibility. Achieved 95%+ verification accuracy on multi-subject test sets.
- Optimized model inference for real-time deployment on embedded ARM hardware with <500ms inference latency.
National University of Singapore
Academic Intern
Carma — Real-time AI-Powered Violence Detection
- Architected end-to-end video classification system combining LRCN and ConvLSTM for temporal modeling of violent events in CCTV streams. Achieved 83% F1-score on custom surveillance dataset.
- Implemented spatial-temporal feature extraction capturing both appearance and motion dynamics under occlusion and illumination variation.
- Deployed full-stack pipeline: Streamlit frontend, AWS Lambda backend, Firebase real-time database with sub-second alert latency and geolocation tagging.
Shiv Nadar Institute of Eminence
Student Researcher & Research Assistant
Accelerated Sparse Diffusion MRI for Neurodegenerative Disease Detection
- Developed Swin-Transformer architecture maintaining diagnostic accuracy with 50%+ fewer gradient directions (5–21 vs. standard 41), enabling clinically feasible acquisition protocols for elderly patients.
- Designed custom loss function with positive-definiteness constraints and domain-specific penalty terms to enforce biophysical coherence in diffusion tensor estimates.
- Conducted voxel-wise TBSS analysis to map disease-specific white matter degradation for tract-level AD and FTD classification without full acquisition. Accepted at ICCV 2023 and ACML 2023.
ARMARecon — GNN for Sparse-Label Classification
- Developed novel GNN integrating ARMA spectral filtering with reconstruction-driven regularisation — the first approach to jointly address over-smoothing and over-squashing in sparse-label neuroimaging graphs.
- Achieved 98.3% accuracy for CN vs. MCI and 99.7% for CN vs. AD on ADNI dataset (>3% above GCN/GAT baselines). Accepted at ISBI 2026, London.
IIIT Hyderabad — CVIT Lab
Research Intern
Facial Recognition Bias Assessment & Mitigation
- Built a 30K-image annotated dataset spanning diverse Indian demographics including geographic, ethnic, and age variations.
- Benchmarked 5 commercial and open-source facial recognition systems using ROC curves, AUC, FAR/FRR. Uncovered 15–20% accuracy gap on underrepresented Indian faces vs. Western-centric benchmarks.
- Proposed targeted few-shot adaptation strategy that recovered 12% of the performance gap. Results presented to CVIT faculty panel.
Shiv Nadar Institute of Eminence
Teaching Assistant — Image Processing (CSD357)
- Supported 60+ students across labs on spatial filtering, edge detection, morphological operations, and frequency-domain analysis.
- Assisted with MATLAB and Python programming assignments, graded lab reports, and held office hours for debugging and conceptual clarification.
Shiv Nadar Institute of Eminence
Teaching Assistant — Mathematical Methods I (MAT103)
- Supported 80+ undergraduate students across Computer Science and Engineering on linear algebra, differential equations, and vector calculus.
- Conducted weekly problem-solving sessions, graded assignments and exams, provided one-on-one mentoring.