Before Amazon: four years at IIT Kanpur.

Object detection and classification in surveillance video

A course project in a team of five: separated moving objects from the background by frame differencing, cut them out of each video, and trained multiclass classifiers on HOG features — SVM, decision trees, random forests, AdaBoost and k-NN — to tell autos, rickshaws, cars, bicycles, motorcycles, pedestrians and number plates apart.

A Docker deployer

Deploys any server code from a configuration — MongoDB as a single node, a replica set or shards — behind a load balancer whose servers can be added or terminated at runtime from the UI.

Skyline queries on MapReduce

Adapted the BNL, SFS and bitmap skyline algorithms to MapReduce and benchmarked them on a Hadoop cluster with three physical worker nodes, partitioning by angle rather than by grid so every mapper gets some skyline points.

A catalogue builder for sellers

On a seller-facing team: a tool that turned raw product data into a validated catalogue feed, with errors shown in the UI and guidance for every attribute.

3-D Tetris

Tetris with a third dimension, in C++ with OpenGL and OpenAL — texture-mapped blocks, ambient lighting and particle effects, sound, and a heads-up display with the score and the next block.

Tetris3D-OpenGL-Game

Number plates on the campus cameras

Found the number plates on moving vehicles in the institute’s surveillance feed. A real-time adaptive Gaussian-mixture background model, with shadow detection, threw away empty frames — run on the open-source BGSLibrary; Viola–Jones classified the vehicles; morphological operations isolated plate regions as the first step towards OCR — tuned to hold up through changes in light and weather.

Image_processing

jutja.com

Co-founded a project-management site that laid tasks out as mind maps, and built its front end — a single-page interface, with the mind-map visualisation in JavaScript on VivaGraph.

A Java-to-MIPS compiler

A cross-compiler written in Python by a team of three: a lexer, a parser that emits parse trees for Graphviz, three-address code generated from semantic rules on the grammar, and a MIPS back end with register allocation and runtime library support.

Compiler-Assignment-1Compiler-Assignment-2Compiler-Assignment-3-and-4

A content-based email classifier

Sorted the public Enron corpus into categories with a bag-of-words model — multiclass SVM, naive Bayes and random forests, with POS tagging and stemming — in Python with NLTK, scikit-learn and Weka, plus a Gmail prototype in Apps Script. Up to 80% accuracy on balanced sets.

Planar graph visualisation

Graph-drawing heuristics for planar graphs, and ways of minimising edge crossings for different kinds of data.

Hexxagon

The hexagonal board game in Python and pygame, for two players or against an AI with easy and hard levels.

Hexxagon