Projects
Systems, models, and tools I've built. Detailed case studies on the engineering decisions, constraints, and results behind each one.
NTIRE 2026 — Efficient Low-Light Image Enhancement
Building an image enhancement model under a strict computational budget.
The challenge was not simply to make an image enhancement model perform better. It was to make it small enough to be useful. Working under a 1 MB model-size constraint forced me to think carefully about architecture, parameter efficiency, and which improvements were actually worth their computational cost.
I worked on a lightweight image enhancement architecture designed around efficient feature extraction and residual learning, with approximately 250K parameters while remaining under 1 MB in model size.
The work ranked 14th among 1,500+ participants and was included in the CVPR 2026 Workshop Proceedings, with me as a co-author.
Results
- Model size<1 MB
- Parameters~250K
- Rank14 / 1,500+
- PublicationCVPR 2026 Workshop Proceedings
Technologies
E-Sannidhi
An offline-first telemedicine platform designed for low-connectivity environments.
Built during the Smart India Internal Hackathon 2025, E-Sannidhi was designed to connect healthcare workers and patients in rural areas where reliable internet connectivity could not be assumed.
The system was built around an offline-first approach: healthcare workers needed to continue capturing patient workflows without a network connection and synchronize their data when connectivity returned. That made local persistence via IndexedDB, synchronization, and conflict handling central design concerns rather than features added after the interface was built.
Our team ranked 31st among 200+ teams.
Key ideas
- Healthcare worker–patient workflows
- Offline-first data capture
- IndexedDB persistence
- Synchronization when connectivity returns
Emotion-Aware Assistant
Exploring how emotion recognition and language models can make conversational systems more context-aware.
This project explored a conversational AI system that uses detected emotional context to influence how responses are generated.
The system combined emotion classification with a language model, using the detected emotional state as additional context rather than treating conversation generation as an isolated language-model problem.
The work involved experimenting with transformer-based emotion modelling, sequence modelling, and conversational response generation, with a focus on understanding how emotional context could be incorporated into an interactive system.
Key ideas
- Emotion-aware conversational responses
- Transformer-based emotion modelling
- Sequence modelling
- Context-aware response generation
CampKart
A full-stack campground marketplace built around listings, authentication, and transactional web interactions.
CampKart is a full-stack marketplace for managing campground listings and user interactions. I built the application across the frontend and backend, connecting a responsive React interface to RESTful APIs and a PostgreSQL database.
The system supports CRUD operations, user authentication, protected routes, reviews, and listing management, giving me experience working across the complete flow from user interaction to backend API and persistent data.