I build AI agents
that remember,
adapt, and disagree.
Third-year CS student working on multi-agent AI — research simulations, an agentic tutoring platform, and a diagnostic tool that explains its own reasoning. I like giving agents memory and behavior so they act less like scripts and more like people.
A bit about me
I'm a third-year Computer Science (AI) student at IGDTUW, Delhi. Right now I'm working as a Computational Social Science Research Intern on PolLab, a platform that uses LLM agents to study online polarization and opinion dynamics.
Most of my projects circle back to the same idea — agents that remember past interactions, form opinions, and don't always agree with each other. That's shown up in a polarization simulator, a grading and tutoring platform, and a diagnostic tool that has to explain its own reasoning.
Outside of that, I spend a lot of time on competitive programming — 550+ problems solved on LeetCode, and I made it to the finals of the PayLoad Hackathon.
What I build with.
Languages
Frameworks
Tools & Platforms
Core
A few things I've built.
PolLab
Live demo ↗A research platform for studying online polarization using LLM-driven multi-agent simulations. It has five optional cognitive layers — Personality, Memory, Emotion, Cognitive Bias, Behavior Controller — running on a Flask + SQLAlchemy backend, with dashboards and surveys for researchers.
MediScan AI
Live demo ↗A chest X-ray classifier for detecting pneumonia, built on EfficientNet-B0 — 84.9% accuracy and 91.8% recall on a 624-image test set. It also generates Grad-CAM heatmaps and turns them into plain-language explanations using Groq's LLaMA 3.3, so it's not just a black-box prediction.
GradeSense
Live demo ↗A grading and tutoring platform I built as my SIP 2026 capstone, leading Team TechTitans. It runs on a custom multi-agent pipeline — plain Python with Pydantic, no LangChain or CrewAI — where separate agents handle grading and tutoring, powered by Groq's Llama models.
Blood & Organ Donor Finder
Demo ↗A full-stack platform for matching blood and organ donors with hospitals, with role-based access for both. Matching uses the Haversine formula to find nearby eligible donors and send automatic alerts, with WHO/NBTC eligibility rules built into the request flow.
Currently.
Computational Social Science Research Intern
- Developing an enhanced CSS research platform supporting LLM-based multi-agent simulations of online polarization, opinion dynamics, and other social behavior phenomena.
- Designing a modular five-layer cognitive architecture — Personality, Memory, Emotion, Cognitive Bias, Behavior Controller — to improve contextual reasoning and behavioral realism in simulated agents.
- Engineering scalable simulation infrastructure with Python, Flask, SQLAlchemy, SQLite, and Groq Llama 3, while researching extensible frameworks for CSS experimentation.
A couple more things.
Co-authored a machine learning paper, published in a Springer-indexed conference (top 20% acceptance rate).
600+ problems solved across core DSA topics on LeetCode. Peak contest rating: 1666.
Reached Round 3 (semifinals) of Flipkart GRiD 8.0 — selected from 1,65,730 registrants nationwide
WIEGNITE 7.0 (IEEE IGDTUW) — Special Mention among 190+ teams for a sustainability-focused simulation game.