Computer Science undergraduate at Thapar Institute of Engineering & Technology. I build machine learning systems end to end — perception, model, evaluation, deployment — and I care more about whether a result survives scrutiny than whether it looks good in a summary table.
Recently completed an AI/ML research internship at C-DAC Noida, building an automated evaluation system for handwritten examination booklets.
EvalDesk — grades handwritten exam booklets end to end. Vision–language OCR (selected by benchmarking six engines), order-independent answer-to-question mapping via the Hungarian algorithm, and an LLM scorer paired with a contradiction-aware verifier that flags uncertain marks for human review instead of failing silently. Mean absolute error of 2.50 marks against human examiners, with no systematic bias. Paper under review.
Parkinson's Voice Detection — a re-analysis — the UCI Parkinson's dataset has 195 recordings from 32 people, so a random train/test split puts the same voice on both sides. I found this in my own coursework submission and measured it: the standard SVM result drops by up to 24 accuracy points under subject-level cross-validation, landing below the majority-class baseline.
Snake AI — Neuroevolution — 1,000 feedforward networks evolve to play Snake with no gradients, no demonstrations, and no hand-coded strategy. Egocentric ray-cast perception, tournament selection, parallel fitness evaluation.
Music Streaming DBMS — relational schema normalised to 3NF, with the decomposition of every 1NF/2NF/3NF violation documented, plus PL/SQL stored functions and analytical queries.
Evaluation methodology for ML systems · document AI and OCR · neuroevolution and gradient-free optimisation · anything where the hard part is deciding what "correct" means.
Away from the terminal: guitar, music production, and badminton.
Python · PyTorch · NumPy · scikit-learn · Hugging Face Transformers · OpenCV · Flask · Docker · SQL · Git