Projects

1. Natural, Expressive, Indian-Accented English TTS

   I built a Text-to-Speech system that generates natural, expressive Indian-accented English speech. It was my foray into the world of TTS, from understanding the nuances of speech synthesis to fine-tuning a model that can handle the intricacies of Indian accents. I deployed the model with streaming support, making it production-ready for conversational use cases, with a RTF of 1.2x on a A10G GPU.

   See the Blog!
2. An LLM Built From Scratch

   I wanted to understand language models not just as a consumer or student, but from the inside. So I built a GPT-2–like model from scratch — everything from tokenization to training loops. Just raw PyTorch and text files. That's when I truly appreciated what goes into making language models tick.

   GitHub
3. Mathematical Ability

   I spent a whole semester deep-diving into a highly mathematical and notation-heavy paper on Strong Rank-Revealing QR Decompositions. I wrote code in Fortran (thanks to autoregression, picking up new stuff is much easier now), managed my team, parsed dense academic papers, and even explored LAPACK's source code. It was painful, dense, and one of the most satisfying technical journeys I've taken — it made me fall even deeper in love with numerical linear algebra and low-level programming.

   See it for yourself
4. Data Visualization and Story-Telling

   Data's only useful if it moves people. Over time, I've built a knack for taking messy, dense datasets and telling stories with them — stories that are clear, sharp, and sometimes even fun. It's not just about charts, it's about narrative flow. See it for yourself:

   My Quintessential Work on Data Viz: Here | Code

   My Take on Famous Burtin's Antibiotics: Here | Code | Poster
5. Tools for Research in Vision-Language Models

   I built a full suite of tools to manage and support ~50 annotators working on creating a benchmark dataset for VLMs, and to run model inference for evaluation on that dataset. It was mission-critical — with little room for failure, since both annotation and inference are expensive — and focused on streamlining the entire pipeline, from dataset creation to evaluation. A small but meaningful step toward the larger vision: a Ten-Trillion-Token corpus for Indic languages.

   GitHub
6. Graphs Visualizer for Adjacency Matrices

   I wanted students of Graph Algorithms (like myself at the time) to see and feel graphs, not just trace through lists and arrays. Figured out SVG through pure documentation-diving — it was pre-GPT era — and built something that made graphs easy to work with. Looking back, it was my first brush with building educational tools that felt useful.

   Use it here!
7. Not Just Code: Leading, Building, Shipping

   This was where theory met application. I led a team to develop and deploy a simple app based on the Nearest-Neighbor algorithm. The application itself wasn't complicated, but the approach is what stands out. I learned to juggle code, bugs, timelines, and people — and somewhere in there, how to ship end-to-end products that not only work, but are beautiful too.

   See Here (Static-Build) | Frontend (Vidhi) | Backend (Merupa)
8. Stadiometer

   Built a functional stadiometer from scratch — the kind that measures your height. This wasn't just code: it involved PCB design, CAD, laser cutting, wood milling, embedded programming, and a whole lot of debugging. It taught me the joy of building things that live outside the screen. Atoms, not bits.

   Documentation
9. Flagstaff

   We've all grown up watching Americans fly their flags in the most aesthetic way possible (think Wall Street). In India, I saw a very different story — cheap fabric, awkward poles. So I built an Indian flagstaff — minimalist, elegant, and fully compliant with the Indian Flag Code. More importantly, I sold them too (4 units at ₹1500 each). It's my little tribute to doing things with grace and intention.

   Blog