A short note on who I am and what I work on.
I studied Mathematics and Scientific Computing at IIT Kanpur, with a minor in Machine Learning. Since then I've worked as an Applied Scientist across credit risk, conversational AI, federated learning, and computer vision — building the models and the training and serving infrastructure around them.
Recently that's meant a corporate bankruptcy-prediction model now used by PayU for real-time risk monitoring, and a document-review agent used by the NSE to cut review timelines from weeks to days. Outside work I read a lot of papers and try to explain them in plain language — if something here is useful or wrong, I'd genuinely like to hear about it.
Projects & things I've built.
Live Transcription System
Real-time speech-to-text over WebSockets — Whisper transcribes live audio at 2–3s end-to-end latency, fanned out to clients via a Redis queue buffer.
Mosaic Single-Cell Integration
A three-encoder–decoder deep generative model that aligns RNA and protein single-cell assays into a joint embedding. With Prof. Hamim Zafar, IIT Kanpur.
Generalised Few-Shot Segmentation
A Local Feature-based Multi-Prototype Classifier that captures diverse class features, improving novel- and base-class accuracy on Pascal-VOC. With Prof. Anil K. Tiwari, IIT Jodhpur.
Where I've worked.
Building applied machine-learning systems end to end at an early-stage startup in Hyderabad.
Built a corporate bankruptcy-prediction model (~0.8 F1) deployed at PayU for real-time risk monitoring, and an AI review agent for DRHP documents used by the NSE — cutting reviews from weeks to 1–2 days.
Built FastAPI services to deploy LLMs on ephemeral compute, plus regex-based feature selection and label-free model deployment.
Revamped a chat-based wellness tracker with multi-intent handling, and built LangChain/LLM data pipelines and a RAG contract-analysis prototype.
Added semantic caching (LangChain, GPTCache, Pinecone) and a Google-Translate/GPT translation hybrid — cutting similar-query response time 75% and translation time 66%.
With Prof. Dusit Niyato — a Clustered-Layer Personalized Federated Learning method beating FedAvg by 35% on non-iid MNIST, plus automated resource allocation across the mobile-edge-cloud continuum.