Building intelligent
systems that matter

A novel framework built on PyTorch for dynamic data mixing during LLM training. Supports static and dynamic sampling strategies with on-the-fly mixture adjustment — integrated with IBM's internal training stack.

A drop-in extension for standard SFT that brings the benefits of few-shot ICL into fine-tuning. Uses the base model as a "free teacher": an 8-shot context produces rich token-level soft labels that supervise a LoRA student alongside cross-entropy loss — capturing dark knowledge and task-specific registers that one-hot labels discard. Evaluated on GSM8K across multiple model families, FSD consistently outperformed standard SFT with gains up to +8.4 pp, with the Qwen3-8B variant surpassing its own 8-shot teacher — all at zero-shot inference speeds.

Trained a VLM to generate code from UI images and PDFs using reinforcement learning with verifiable rewards — without SFT. One of the first works exploring RLVR for image-to-code tasks.

Implemented online data mixing for pre-training BharatGen models, eliminating the ablation costs of iterating over different static mixtures — saving hundreds of hours of compute while delivering best-in-class performance.

Built a document image translation pipeline with a novel layout detection module that models layout as a token classification problem, grouping related tokens into segments for better translation.

Developed an algorithm using submodular functions to select the optimal data subset yielding highest returns. The resulting models consistently outperformed random subset baselines on downstream tasks.

Built a RAG-based framework that retrieves relevant legal documents for each query, making chatbot responses more accurate and explainable. A joint initiative of IIT Bombay, NLU Bangalore, and Meta.

Combined predictions from TrOCR (vision) and RoBERTa (language) through an innovative multi-model logits fusion approach, significantly improving OCR accuracy on handwritten text.

A compact autograd engine implementing backpropagation (reverse-mode autodiff) over a dynamically constructed DAG. Includes a small-scale neural networks library with a PyTorch-like API.

Real-time people counting system using OpenCV with SSD + MobileNet for detection and a centroid tracker for tracking. Features alert systems for occupancy compliance and threaded video processing.

Automated attendance system for online classes using a Selenium web scraping bot, React admin panel, and Flutter mobile app. Received a special appreciation award from the college.

Built the cultural fest website and an offline automation system to streamline ticketing for 2000+ participants with ID card scanning to prevent fraudulent re-entry.