I'm Siddharth Padmakumar, sidp for short. I'm a master's student in Artificial Intelligence and Innovation at Carnegie Mellon. Before this I spent two years as a software developer at IBM, shipped Bill2Go to Google Play, and co-authored a paper on anomaly detection in spatio-temporal graphs.
MSAII at Carnegie Mellon, Pittsburgh. Class of 2028.
~/for-the-vibes
Things I built because I wanted them to exist. Nobody assigned these. Some stayed small, and one ended up on the Play Store.
Scan any receipt, let a vision model itemize it, share a no-login link, and everyone claims what they ate in real time. Twenty minutes of post-meal math becomes under one minute.
// Splitting a restaurant bill is a small annoyance that everyone has. I wanted each person to pick their own items from a shared link instead of one person doing the math.
React Native · Expo · FastAPI · Firebase · Gemini Vision · TypeScript · RevenueCat
open project →Practise dance from any YouTube video: mirror it, slow it down, loop a section, then split the screen with your webcam and record yourself dancing alongside the reference, in sync.
// A practice tool for learning dance from YouTube. Mirror the video, slow it down, loop one section, and dance next to it on your webcam.
Next.js 15 · React 19 · TypeScript · Tailwind · NestJS · Supabase · Firebase Auth · pnpm
open project →~/the-official-record
College, internships, and published work. Most of it was done in teams, and each page says who was on them. My IBM work was internal, so it lives on the about page as cards instead of repos.
Peer-reviewed method for finding anomalies in time-varying graphs: predict each sensor's normal future with an attention-based graph network, then localize what's misbehaving from subgraph down to single node.
Python · PyTorch · PyTorch Geometric Temporal · A3T-GCN · SciPy · METR-LA
open project →A faithful simulation of India's UPI instant-payment system: a client and three cooperating Spring Boot servers (Bank, NPCI, UPI) handling registration, PIN-authenticated transfers and per-transaction logging end to end.
Java · Spring Boot · Spring Data JPA · WebFlux · Thymeleaf · MySQL · Maven
open project →Change the emotion of a voice recording (angry to sad, and back) without changing the words: speech becomes a mel-spectrogram image, a CycleGAN translates it between emotion domains, and Griffin-Lim turns it back into audio.
Python · TensorFlow · Keras · CycleGAN · librosa · Google Colab
open project →~/exploring
Two ideas I'm working through: making a shared kitchen easier to manage, and helping coding assistants pick up where another left off.
Feed it fridge photos, grocery receipts, each person's goals and the recipes we already know how to cook. It would keep one running inventory of the kitchen, then produce meal plans for the whole apartment and a shopping list for the gaps. The part I expect to be hard is matching and deduplication: deciding that the milk on a receipt and the carton in a photo are the same thing.
planning to use: CV and OCR · learned matching · retrieval and ranking · LangGraph · a constraint solver
Coding assistants lose their context between sessions, laptops and teammates. The idea is to capture it through the hooks the assistants already support, then hand off selected parts with permissions respected, with decisions, provenance and Git state versioned alongside. Agents could use the same layer to coordinate work. I want to test it against the obvious baseline, shared Markdown files, and against Acontext.
planning to use: assistant hooks · permission-aware handoffs · versioned decisions and provenance
I'm also learning how transformers and neural-net training work inside, when to fine-tune instead of using RAG, and how to connect computer vision and LLMs end to end.
Smaller experiments and coursework. No pages, just the record.
