M.Tech in Computer Science & Engineering at IIT Kanpur. I work at the intersection of deep learning, privacy-preserving computation, and scientific visualization — building calibrated, secure, and interpretable systems.
I'm a graduate researcher at IIT Kanpur, drawn to problems where reliability, privacy, and interpretability matter — especially in safety-critical vision and data-sensitive systems.
My recent work extends Deformable DETR with aleatoric & epistemic uncertainty, reducing calibration error by 72% while preserving mAP. Alongside, I build privacy-preserving pipelines using secret sharing, MPC, and PIR, and parallel scientific-visualization tools with MPI + VTK.
I've served as Teaching Assistant for 4 courses at IIT Kanpur — from first-year programming to DevOps and databases — and I enjoy mentoring, visualizing data, and shipping end-to-end prototypes.
Indian Institute of Technology, Kanpur
GRD Institute of Management & Technology, Dehradun
Extended Deformable DETR with aleatoric variance prediction, MC dropout (T=15) for epistemic uncertainty, and uncertainty-weighted fine-tuning. Reduced Expected Calibration Error by 72% (0.128 → 0.035) on COCO val2017 while preserving mAP (43.3 vs 43.8) and improving APS.
Hybrid pipeline combining text embeddings + LSTM, PCA-reduced deep features, and SVM ensemble with a Random Forest meta-learner.
Phishing detector using 21 URL/HTML/JS/traffic features augmented with Naïve-Bayes probability. Random Forest hit 96% accuracy with ~3.6% false positives.
End-to-end resume ↔ JD scorer parsing PDFs, extracting entities via NLP, and ranking with semantic similarity. Reached ~0.70 correlation with human ratings.
Prototype regularization + generative replay on CIFAR-10 with Gaussian prototype classifier; PCA pipeline for incremental updates.
Linear-regression pipeline for absenteeism with group-wise fairness auditing. Bias mitigation achieved 0.00 MAE/RMSE/prediction gap across age, education, and service-time groups.
Matrix-factorization recommender with two-party additive secret sharing. Uses PIR to securely fetch item features and update secret-shared user profiles — simulating secure multi-party computation end-to-end.
4-party secret-shared multiplication over TCP sockets using Beaver triples. Implemented Basic, Sequential, and Parallel protocol variants with a triple generator and share distribution.
Privacy-preserving retrieval via PIR with Additive and Shamir's Secret Sharing. Client-server protocols benchmark query creation, reconstruction time, and bandwidth across schemes.
Threshold-based secret splitting and reconstruction using polynomial interpolation — validated recovery from any t shares.
Interactive Streamlit dashboard analyzing global vaccination trends, manufacturer dynamics, and correlations with health outcomes. Maps, scatter, line, bar, and pie views over cleaned, merged datasets.
Distributed ray-casting with early termination and sub-image compositing. Scalability / timing analysis on 1000×1000×200 and 2000×2000×400 volumes.
MPI-parallelized renderer with 1D slicing + 2D block decomposition, colour/opacity transfer functions, and final composited output.
Isocontour extraction from raw grids exported to ParaView, plus VTK volume rendering with vtkSmartVolumeMapper, custom transfer functions, and Phong shading.
Jupyter tool linking isosurface rendering with live histogram updates through an isovalue slider, plus reset & reproducible UI controls.
Particle path tracing in 3D flow fields using RK4 integration — forward and backward streamlines from seed points with vector interpolation.
HCI design study: identified 45% urgent-ride cancellations as a pain-point and proposed keyword-based detection, priority tags for drivers, and restricted cancellations. Built functional UI prototypes for both customer and driver flows.
Camera controller for photo / HDR / video via hand gestures and voice commands, plus music playback controls — running on a Raspberry Pi with ultrasonic sensor, Mediapipe, and speech recognition.
Posture classifier (sitting / standing / lying) from PhonePi+ accelerometer data using threshold-based analysis and visualization.
Simulated GPS localization from satellite signal travel times; analyzed timing-error impact with visualized error trends.
Interactive visualizer for Selection, Bubble, Merge, Quick, and Insertion sort — animated states for comparing, swapping, sorted, and pivot, with speed/size sliders and play controls.
Designed & evaluated an OOP-in-C++ assignment on memory-efficient file indexing. Reviewed 230+ submissions, assisted exam invigilation and quiz grading.
Supported 120+ students across grading, project evaluation, and invigilation. Mentored 3 groups on research-paper evaluation projects.
Grading, malpractice review, and course-administration support.
Mentored first-year undergraduates in a cohort of 600+ students/semester — doubts, grading, and malpractice checks.
Part of a 250+ member guide team helping new PG students with academic and social transition at IIT Kanpur.
Explained how query-driven methods prioritize user-defined subsets for targeted visualization. Analyzed the DEX framework's bitmap indexing for fast, scalable exploration of large datasets.
How Banner enables seamless resolution switching in image sensors — framework architecture balancing resolution, bandwidth, and energy for mobile/embedded vision.
Open to research collaborations, internships, and interesting conversations. The fastest way to reach me is email.
