Projects
Published research, coursework I took further than the assignment asked, and a couple of projects that existed purely because the idea wouldn’t leave me alone. Newest first.
2026 · AAAI
The question was whether a buyer’s transaction history and their position in the buyer–seller graph say the same thing about delivery-refusal risk. They don’t: a transformer over sequences and topological signal from the interaction graph turn out to be complementary. Co-authored, and the model scores every order placed on Meesho at checkout.
2022 · M.Tech. thesis
Monocular SLAM built on predicted depth inherits every mistake the depth network makes. So I fed per-pixel uncertainty into the pipeline as a first-class input rather than a post-hoc score, added a temporal-consistency objective across the video, and trained the depth model with no ground truth at all — by backpropagating an ORB-SLAM trajectory signal through gradSLAM.
2021 · Information Retrieval
Playlist continuation from a seed set of tracks. Neural collaborative filtering against a well-tuned nearest-neighbour baseline — and the baseline won, which is the more useful result. NDCG@500 of 0.215 versus 0.332.
Prof. Srikanta Bedathur, IIT Delhi.
2020 · Information Retrieval
An end-to-end tf–idf retrieval system supporting prefix wildcards
(mit*) and typed entity queries
(L:new* → documents mentioning New York). Replacing the linear
posting-list scans with a hash-based index cut construction time 17%.
2020 · Independent
Co-watching for YouTube and Vimeo, built on Node.js socket programming: an event-driven channel between browser and server keeps playback state consistent across everyone in a room. Independent rooms, live chat, and a shared ambient-lighting control, because why not.
2020 · Independent
An excuse to take one model the whole distance: semantic segmentation and facial feature recognition behind a Flask app, containerised and deployed. The part worth learning was never the model.
2019 · Computer Vision
Recovered camera intrinsics from multiple marker views, solved for the homography from marker pose, and augmented 3D objects into the scene. Then extended it into a two-marker AR ping-pong game driven by real-time pose estimation across both marker planes.
2020 · Machine Learning
Decision trees built from scratch — gini-index splits, then pruning — before handing the same problem to grid-searched random forests to see how much the ensemble was really worth. 0.807 accuracy on the UCI dataset.
Also
Qualification round. Top 0.5% nationally in JEE Mains 2017 and top 2% in JEE Advanced, out of 1.1 million candidates — the last time a percentile meant anything to me.