Software Development Engineer at Zupee · B.Tech, Electronics and VLSI Design, IIIT-Delhi (2022–2026)
I work on backend systems and machine learning. At work I build TypeScript / Node.js services; outside work I build PyTorch, computer-vision and LLM-agent projects, and I try to report their results honestly, caveats included.
- Now: SDE at Zupee (Gurugram). Event-driven APIs (Kafka, SQS, WebSockets), payment and KYC integrations, and React / Node.js microservices on MongoDB, Redis and Kubernetes.
- Before:
- AI Intern at Mstack Chemicals: agentic LLM workflows and retrieval-augmented Bayesian optimisation for chemical reactions.
- Undergraduate researcher at CAI Lab, IIIT-Delhi: memory-efficient training for ultra-high-resolution object counting.
Machine learning and research
| Project | What it is |
|---|---|
| countgd-patch | Patch-based, fractional-batch training that fits CountGD (open-world object counting) on an 11 GB GPU. With 4 scales and FPN fusion it stays within 0.5 MAE of the full model. |
| fsc147-vlm-annotation-refinement | Finds missing labels in FSC-147. 372k model predictions go through a geometric filter, a Gemini 2.5 judge and a Streamlit human review, which add 566 verified points across 161 images. |
| counterfactual-llm-popularity-bias | Counterfactual prompting of Qwen2.5, anchored on ItemKNN, to reduce popularity bias in MovieLens-1M recommendations. |
| meta-nca-1d-arc | A hypernetwork generates a neural cellular automaton for each task, for few-shot reasoning on 1D-ARC. Includes an audit of the evaluation. |
| meta-learning-nifty50-trading | A Meta-SGD + SAM Transformer that adapts online to market regimes, giving Nifty 50 long/short signals that are backtested under transaction costs. |
| Multimodal-Item-Price-Predictor | Amazon ML Challenge 2025: predicts product prices from text and images with frozen CLIP, EVA-02 and DeBERTa encoders plus fusion heads. |
Software and systems
| Project | What it is |
|---|---|
| pairwatch | A Claude Code mod in TypeScript that explains an agent session as it runs, with a live stage, short narration and about 160 concept cards. |
| code-twin | A two-agent TDD loop. A QA agent writes pytest tests, and a dev agent writes code until they pass in a no-network Docker sandbox. |
| edgeforge-ai | A FastAPI service that turns one layout into labelled synthetic computer-vision data, using SDXL + ControlNet, tiled VAE decoding and YOLOv8 auto-labels. |
Languages: Python, TypeScript, JavaScript, C++, Java, SQL, C
ML: PyTorch, scikit-learn, Hugging Face Transformers, Diffusers, YOLOv8
Backend and infra: Node.js, FastAPI, React, Kafka, MongoDB, Redis, Docker, Kubernetes, AWS
- Amazon HackOn 2025: top 130 of 52,000+ participants with TrustSight, a fraud-detection system that uses multi-task learning and graph neural networks.
- Selected for Amazon ML Summer School in 2024 and 2025.
- Codeforces Specialist (1400+); 500+ LeetCode problems solved.