AI & Data Science student at IIT Jodhpur — I build across ML research, systems, and full-stack product, and care about getting the details right.
I'm Agam, a BTech student in AI & Data Science at IIT Jodhpur. I like building software end to end — from research prototypes to full-stack products — and I care about the details that make systems fast, reliable, and genuinely useful.
My work spans machine learning, high-performance computing, and full-stack development. From analysing lunar XRF data for ISRO to stress-testing frontier LLMs and publishing research on dynamic-graph algorithms, I'm drawn to problems where the engineering matters as much as the result.
Outside of work, I'm a big Formula 1 fan and a keen gamer — I enjoy the strategy and engineering behind the sport, and a little of that personality shows up in how this site is built.
Internships and contracts across AI tooling, full-stack product, and quant systems — most recent first.
Prototyping agentic-AI tooling for cloud-native development — LLM-driven agents that automate and streamline infrastructure workflows across the Kubernetes team.
Engineered adversarial test suites that surface failure modes in frontier LLMs (incl. Claude Opus 4.8) on real-world SWE tasks, and benchmarked AI solutions to ML-competition problems against structured rubrics.
Architected the full stack — NestJS backend + React Native (Expo) app — with JWT / Apple / Google sign-in, SecureStore, OTP login, and a SaarthiCoin wallet for a clean cross-screen experience.
Built a TypeScript engine that backtests historical NIFTY data via the Zerodha Kite API and streams realtime quotes over WebSockets, plus an MCP server exposing live market data to downstream AI apps.
From front-wing front-end to rear-axle backend — the toolkit I run when I'm building, training, and shipping.
Published & in-review research in high-performance computing and dynamic graph algorithms. SPADE Lab · IIT Jodhpur · Prof. Dip Sankar Banerjee.
National-level competitions and the teams I've led where the work cleared the gap.
3RD3RD
LEADLEADSelected projects spanning AI research, full-stack product, and a podium-finishing space mission.
Diagnosed four compounding bugs in a Vision-GNN pipeline — including a silent no-op that disabled all graph convolutions — then redesigned the stem with learnable soft-threshold graph connectivity, SE attention and dual max+mean aggregation for a major top-1 lift on CIFAR-100.
A cross-modal heterogeneous graph neural network for weakly-supervised crime anomaly detection on UCF-Crime — grounding DINOv2 visual entities against RoBERTa-extracted report entities through a GATv2 + Heterogeneous Graph Transformer pipeline with a multiple-instance-learning ranking objective.
A modern e-commerce platform with an intelligent recommendation engine based on purchase history and recently-viewed items — cart management, Google OAuth, and a fully responsive Next.js front-end.
An image-classification system on the EfficientNetB0 backbone with transfer learning, distinguishing real from manipulated media. Took the podium against a national field at Predictathon.
A data-analysis platform for ISRO's Chandrayaan-2 XRF spectral data — Kriging interpolation, 3D lunar surface mapping, and detection of Calcium & Titanium signatures. Podium across all 23 IITs at Inter IIT Tech Meet 13.0.
Open to internships, research collaborations, and freelance builds — or a conversation about either Hamilton's championship count or your next ML pipeline.