Projects & competitions.
Most of these started as competitions, the deadline pressure suits me, and a few as self-directed builds. Together they're where I learned the craft that later carried into production work: modelling under class imbalance, validating without fooling yourself, and shipping something that runs.
Competitions
Credit Score Algorithm - 1st place of 235 teams, 800+ participants
- Developed 5 models · ensembles and deep learning, to identify credit defaults across 3,000+ borrowers using 28 parameters, spotting a 20% fraud rate at 92% accuracy.
- Built the scoring mechanism from banking first principles, PoD, LGD, and EAD · and stress-tested a bank model that generated $22,400 max profit at a 789-score lending threshold.
Log Anomaly Detection · AIR 13 of 213 teams, 2nd among IIT Bombay
- Categorised normal vs. abnormal logs at 99.99% accuracy via event categorization and pattern learning.
- Tamed 4.1M+ logs with an 8:500 abnormal-to-majority class imbalance, clustering with NLP, then pruning redundant clusters to isolate the true log labels.
Oil Price Forecasting with Dynamic ML · Top 6 among 40+ final submissions
- EDA on WTI crude established dependence on 3-day and 9-day moving averages; predicted prices via regression and a dual moving-average crossover strategy.
- Forecasted two months ahead with a CNN-LSTM under walk-forward validation - 78% accuracy · and mapped world events against the variance in predictions.
Robot 1.0 · Top 5, GameDev Hackathon by DevCom IIT Bombay
- A 2D browser game built with HTML, CSS, JavaScript and the Phaser framework, still live, still laptop-only, still fun. The top-5 finish came with a direct interview offer from IIT Bombay's developer community.
Machine learning & data builds
Introduction to Machine Intelligence · Seasons of Code, WnCC
- Implemented a handwritten-digit classifier (CNN + TensorFlow on MNIST), then moved to reinforcement learning, Q-learning for optimal pathing.
- Built a Deep Q Network agent from scratch, able to traverse slippery, stochastic terrains and mazes, the kind of agent you'd deploy for route decisions in hard-to-navigate environments.
WorldQuant University, Applied Data Science Lab (three self projects)
- A/B testing: designed an email-campaign experiment over 5,025 students' data (MongoDB, choropleth mapping), estimated the group size of 196 needed to separate signal from noise, and tested significance with a chi-square test at p = 0.05.
- Volatility forecasting (India): built an SQLRepository class for structured market data, fit a GARCH model with 3-day lags set via correlograms, assessed on AIC/BIC and standardized residuals.
- Customer segmentation (US): compared 351 characteristics across 28k+ consumers, selected top-5 features by trimmed variance, and tuned K-Means to k=4 at a 0.693 silhouette score.
Research projects
COVID's impact on construction-industry finances
- Analysed 8+ financial ratios across 10 construction firms, revealing a 15% post-COVID increase in profitability and liquidity, alongside a strategic shift to debt reduction and asset-based financing despite a 12% fall in activity ratios, driven by contract renegotiations and supply-chain disruptions.
Economics of money, banking & financial markets
- Assessed listed coupon bonds (yield vs. maturity vs. credit rating), modelled the term structure of interest rates from T-bills and G-secs, and evaluated the M3 money supply vs. GDP growth relationship over 20 years.
For the non-technical side, sports, music, and the odd student-parliament stint, see extracurriculars & achievements on the home page.