My Projects
Mycroft: Effective and Efficient External Data Augmentation
Thesis project (Under Review at TMLR | Advisor: Prof. Arjun Bhagoji). Collaborative ML requires identifying useful data from multiple owners without full data sharing — but existing approaches either share too much or select too little. Mycroft proposes Adaptive, Agnostic, and Voting-based selection methods that estimate the utility of an entire dataset from just 32 samples per class, and demonstrates robust data-owner ranking even under noisy data conditions.
ChronoSSM: Temporally Aware Autoregressive State Space Models
Research project (Under Review at KDD'27 | Advisors: Prof. Arjun Bhagoji & Prof. Francesco Bronzino, ENS Lyon). Most autoregressive models predict what happens next, but treat when as an afterthought — typically via a frozen, decoupled timing module. ChronoSSM trains a shared backbone to jointly optimize event prediction and inter-arrival timestamp prediction, letting temporal gradients directly reshape representations. Across 4 domains (business processes, clinical records, network traffic, knowledge graphs), joint training consistently improves temporal recoverability without degrading event generation quality — improving R² from 0.19 to 0.39.
RepAlign: Representation Alignment for Adversarial Training
Research project (To be submitted to ICLR'27 | Advisor: Prof. Arjun Bhagoji). Standard adversarial training methods like TRADES align only model outputs. RepAlign proposes aligning intermediate feature representations instead, using similarity metrics — CKA, CKNNA, HSIC, and Soft-kNN — to bring clean and adversarial representations closer together at every layer. On CIFAR-10, this improves clean accuracy from 83.34% → 86.27% while simultaneously increasing robust accuracy from 52.82% → 54.18%.
Robust 2D Direct Preference Optimization
Course project — CS 769: Optimization for Machine Learning (Guide: Prof. Ganesh Ramakrishnan). DPO aligns LLMs with human preferences but treats all segments of a response equally — ignoring that even "good" responses have weaker parts. 2D-DPO addresses this with finer-grained, segment-level scoring. This work identifies that 2D-DPO is still vulnerable to label/score noise and proposes a robust formulation with segment-level noise robustness, backed by theoretical analysis and empirical verification on a Pythia 6.9B model.
CandleDiff: Contextual Diffusion for Structurally Valid OHLC Generation
Research project (Under Review at ICAIF | Motilal Oswal Financial Services). Existing financial generators only produce return series — not full OHLC bars — and nothing forces outputs to be geometrically valid (H ≥ max(O,C), L ≤ min(O,C)). CandleDiff introduces a candle reparameterization (return, body, upper wick, lower wick) that makes structurally invalid bars mathematically impossible, while conditioning generation on firm fundamentals, a return/volatility scenario, the real market-index path, and recent trading history. Achieves 100% structural validity vs. 71–80% for baselines, and reduces strategy-selection overfitting gaps by 29% in backtesting experiments across 7,280 held-out firm-quarters.
Zero-Shot or Fine-Tuned? Benchmarking Time-Series Foundation Models Across Indian and US Equity Markets
Research project (Under Review at ICAIF | Motilal Oswal Financial Services). Systematically benchmarks 6 time-series foundation models (Chronos-2, Moirai, Kronos, Toto, Falcon-TST, TS-ICL) across zero-shot, last-block, and full fine-tuning regimes on Indian and US equities at daily and 1-minute frequencies — 72 experiments total. Key findings: zero-shot beats fine-tuning in 12 of 24 model–dataset pairs; fine-tuning can cause up to 10× degradation in some cases (Toto on US Daily); and lower price-level error (MAE) does not reliably correspond to better directional forecasting skill.
Unified Multi-Task Image Denoising and Deblurring
Course project — GNR 650: Advanced Deep Learning for Image Analysis (Guide: Prof. Biplab Banerjee). Trained a single model for joint image denoising and deblurring using DINOv3 and LoRA. Demonstrated that multi-task learning generalizes better than task-specific models, improving face deblurring by 0.15 dB PSNR while maintaining denoising performance.
Emotional Analysis for Indic Languages
Course project — CS 626: Natural Language Processing (Guide: Prof. Pushpak Bhattacharya). Trained and fine-tuned transformers on multi-lingual inputs, especially Hindi. Analysed emotion over 5 dimensions using the BHAAV dataset. Compared BERT, LSTM, and CNN models against modern LLMs, achieving 95.32% accuracy with BERT.
CLIP Implementation from Scratch
Course project — CS 725: Foundations of Machine Learning (Guide: Prof. Preethi Jyothi). Implemented the CLIP architecture in PyTorch using ViT and DistilBERT encoders with symmetric contrastive loss to learn shared vision-language representations. Evaluated on Flickr30k for bidirectional image-text retrieval.