Research Blogs

I made my internal deadline tracker public. Search 250M papers, track 12K venues and 2K conferences across 53 fields, and ask in plain language. Over 1,000 researchers on board, built and maintained solo.

SAE experiments tied to a KD paper: toy validation of the minimum-width theorem plus three Pythia-410M trials.

Four protein experiments on one A100: a 1 µs chignolin fold, titin I27 force-pulling, an 8-protein ESMFold sweep in 12.9 s, and ESMFold confidence calibration. Each validated against the published number.

An ongoing investigation into using grokking for better generalisation on under-sampled image data. A first check of off-recipe behaviour on Camelyon17, before the full domain experiments.

Knowledge distillation, mechanistic interpretability, and world models. KD paper under review at COLM 2026.

Circuit-level studies of transformer internals: attention patterns, probing, and activation patching to understand what models actually compute.

An 18-architecture benchmark on PCOS ultrasound. A hybrid CNN-ViT hit 99.81%. A three-stage dedup pipeline cleaned 70% of the public dataset.

Fine-tuning Qwen 0.5B-7B on Indic languages; deploying Gemma 3 1B on Jetson Nano. Compression and diffusion-LM threads now live.

An Agentic RAG workflow with LangGraph and Streamlit. The system prioritizes PDF Q&A and falls back to web search when necessary.

Experiments with Stable Diffusion v1.5 via diffusers: negative prompts, guidance scales, and prompt-engineering basics.

Met Dawar and Dhruv at the India AI Impact Summit and co-started Erdős AI Lab: knowledge distillation, world models, frontier AI. EF-backed, incubated at IIT Bombay.

Representing DSU at India's AI conference in New Delhi: LLM, medical AI, and autonomous drone projects on stage.

The Bangalore Drone Community Meetup: startups, regulations, and the future of UAVs.

A Build-with-AI workshop (Google + Deep Tech Stars) on the Agent Development Kit and the Model Context Protocol.

A tool for comparing ground observations with Landsat Surface Reflectance data and alerting users to satellite overpasses.

A curated collection of 30 research papers that Ilya Sutskever claims contain 90% of what matters today in AI.

A reproduction of Andrej Karpathy's Hacker's Guide to Neural Networks: a code-first approach to backpropagation and real-valued circuits.

Building agentic systems (LangChain, LangGraph, RAG) on internal engineering data at Moog India Technology Centre, Bengaluru.

A practitioner-oriented learning path through large language models: what to read, what to build, what to skip.

Deep learning for medical imaging: automated PCOS detection plus generative augmentation.

A browser-based tool that uses computer vision to find slide transitions in videos and Gemini to generate notes and quizzes.

A teleoperated robotic arm for humanoid robotics: intuitive, low-latency human-robot control.

Building autonomous surveillance drones with real-time computer vision and navigation. More coming soon.