Projects

Production AI products, ongoing research and hackathon builds, from multi-tenant SaaS to deep learning and computational social science.

Next.js, FastAPI, PostgreSQL, pgvector, AWS, RAG, LLMs

Talking Bird

A multi-tenant AI assistant that lets organisations ask plain-English questions across their documents, databases and inboxes, and get cited, calculated answers. An LLM router sends each query to hybrid RAG, an agentic Pandas engine, or both. It connects to sources like SharePoint, Jira, Notion, S3, Gmail and Outlook, drafts email replies from the inbox, and enforces tenant isolation with SSO and per-user data scopes. A licensed build runs at Plaksha University's Office of Research.

Python, Statistical Modelling, FastAPI, React, TypeScript

Clean Air

Tyre degradation intelligence and race strategy console for Formula 1, built for the TrackShift Innovation Challenge 2026 with TGR Haas F1 Team, where it was a top 10 finalist among 1,600+ teams. It removes fuel, track evolution and weather effects from 142,737 laps over 5 seasons, producing intervals 4.4x tighter than the only published model. It enumerates 131,053 strategies per race, and 10 of 11 strategy calls matched what teams actually ran.

PyTorch, GANs, Transformers, Computer Vision, NLP

CrossStyloGAN

A handwriting autocomplete system that reads your handwriting, predicts the next words, and writes them in your own style. It chains a CRNN OCR model, GPT-2 next-word prediction and a transformer-enhanced GAN that captures 657 writing styles. On the IAM benchmark it reached an FID of 4.36 (63% better than HiGAN+) and a CER of 0.65%, at about 1.2 seconds per word on consumer hardware.

FastAPI, Next.js, PostgreSQL, Redis, LLM Infrastructure

Promptly

A drop-in, OpenAI-compatible proxy that cuts LLM spend. Teams point their app at Promptly and keep the same API, while it compresses prompts, prunes context, serves repeats from a two-layer semantic cache and routes requests across OpenAI, Anthropic and Google. Includes per-member API keys, role-based access, savings analytics, spend alerts, and Python and Node SDKs.

Echo Chambers in Israel-Hamas Discourse
Research · Computational Social Science, NLP, LLMs

Echo Chambers in Israel-Hamas Discourse

First-author research on whether online discussion of the Israel-Hamas war forms echo chambers across Reddit and YouTube. Combines LLM-based stance detection, sentiment and network analysis over more than a million comments. Paper in preparation for a conference; findings will be shared after publication.

Ministerial Responsiveness in the Lok Sabha
Research · Political Science, NLP, LLMs

Ministerial Responsiveness in the Lok Sabha

Research on how directly ministers answer questions in India's Lok Sabha, built on a large corpus of parliamentary question-answer pairs spanning more than two decades and measured with embeddings, LLM labelling and human validation. Paper in preparation for a conference; details will be shared after publication.

PyTorch, Transformers, Adversarial NLP

Prism

A research framework testing how well AI-generated text detectors hold up against adversarial attacks. I built the evaluation harness covering 6 detectors (RoBERTa, Binoculars, Fast-DetectGPT, DetectGPT-style, watermarking and a statistical baseline), reproducing the published AUROC scores within about 0.01. The team's character-level attacks evaded detectors 98-100% of the time while preserving meaning.

FastAPI, Next.js, Supabase, LLMs

Synapse

An always-on market intelligence agent for pharma brand teams, and a finalist at MediThon 2025. It polls openFDA, news and ClinicalTrials.gov every few minutes, clusters competitor launches, label changes and recalls, and auto-writes strategic briefs (SWOT, Porter's Five Forces, go-to-market) on a live dashboard.

Python, Physics Simulation, FastAPI, Next.js

SimPulse

A physics-based Formula E race simulator, and a top 5 finalist among 700+ teams at the TrackShift Innovation Challenge 2025. A 100 Hz engine models vehicle dynamics, energy management, tyre wear, overtaking, attack mode, weather and safety cars with stochastic noise, streamed live to a web dashboard and exposed as a reinforcement learning environment.

Python, XGBoost, LightGBM, LSTM, Time Series

Storm Shield

Predicts storm-induced power outages 1-2 hours ahead. A three-stage pipeline forecasts storm occurrence, classifies severity, then routes to a severity-specific outage model, trained on 10 years of hourly NOAA, EAGLE-I and PRISM data. It reaches F1 scores of 0.97-0.99 on outage prediction and 0.85 end to end.

Next.js, TypeScript, Sanity, NextAuth, PhonePe

InstaStay

A full booking platform for guesthouses, live for a real guest house in Barsana. Guests book rooms and halls with PhonePe payments, group bookings, add-ons, refunds and reviews, while admins manage everything from a Sanity CMS. Cron jobs restore room availability automatically after checkout.

Python, Reinforcement Learning, Simulation

Optima RL

Reproduces and extends an AAMAS 2021 paper on the hold-or-not-hold problem: whether a connecting service should wait for a delayed arrival. Trains A2C, DQN, Actor-Critic and DDPG agents across airline connections, a 10-hub freight network and cloud job scheduling, calibrated to FAF5, Alibaba and Google Borg traces, and compares them against fixed-rule baselines.

Sahaay
React, TypeScript, Supabase, i18n

Sahaay

A legal aid platform that helps people fix document errors without paying middlemen. Simple cases go to certified field agents, legal ones to verified low-cost lawyers. Grounded in 15 user interviews, 3 expert interviews and NGO partners, then piloted with 5 users. Built with OTP sign-in, multilingual support and separate dashboards for users, lawyers and admins.