Exploring how intelligent systems are built.
These days, I'm focused on building with LLMs - working on agentic systems, memory, developer tooling, and the designs that turns models into reliable software. I'm particularly drawn to the engineering challenges that emerge when intelligence becomes part of the software stack.
I'm a final-year AI & Data Science student at IIT Jodhpur with a strong interest in the engineering behind intelligent software. What began as curiosity about machine learning has gradually evolved into a broader fascination with the systems, abstractions, and design principles that make modern AI possible.
I enjoy connecting ideas across LLMs, distributed systems, databases, and software architecture. Whether I'm building projects, contributing to open source, or exploring a new topic, I'm motivated by understanding how things work before thinking about how they can be improved.
Architected and developed Oracle Database integrations for modern AI frameworks, enabling OracleDB as a native backend for LLM-powered applications. Built providers for Mastra, LangGraph, and Google ADK, supporting agent memory, workflows, retrieval, and persistence while contributing to Oracle's open-source AI ecosystem.
Conducted research on controllable video weather translation using deep generative models, focusing on transforming scenes across weather conditions while preserving temporal consistency, motion, and visual realism.
Developed customer-facing product features across the full stack using Next.js and modern backend technologies. Built scalable APIs, integrated AI-powered voice capabilities, and shipped production-ready features in a fast-paced startup environment.
Exploring how small vision-language models can recover large-model capabilities by learning from high-quality, filtered supervision instead of larger datasets. Current work focuses on trust-aware pseudo-label filtering, hidden-state based verification, and improving data quality for efficient model adaptation.
A few things I've enjoyed building.
A project I've been thinking about for quite some time. Still designing the architecture, experimenting with ideas, and building it the right way. More soon.
A full-stack AI newsletter platform that researches topics, generates curated newsletters, and delivers personalized editions to subscribers. Designed with scheduled workflows, user authentication, content persistence, and automated delivery in mind, with room for future personalization and recommendation pipelines.
A browser extension that brings contextual AI assistance to any webpage. Askify indexes page content, maintains conversation history, and uses retrieval-augmented generation to answer questions without requiring users to leave the page.
Explored Segment Anything Model 2 (SAM2) for promptable video segmentation, tracking object masks consistently across frames while preserving temporal coherence. Built as an exploration of modern foundation models for video understanding and downstream vision applications.
An interactive visualization platform for machine learning algorithms that makes model behavior easier to understand through real-time parameter tuning and live decision-boundary updates. Designed as a teaching tool and proposed for use in IIT Jodhpur's Machine Learning course.
Improved an open-source vision-language model for automated medical report generation through architectural modifications and inference optimization. Achieved stronger evaluation scores while reducing memory usage enough to enable near real-time inference.