Projects
This is the long-form version of my work—systems, metrics, and links in one place. Featured builds first; older research and hackathon repos underneath. All public code lives on GitHub.
Featured
Clause.ai
- ▸Parent-document RAG on Supabase pgvector—searchable chunks plus full-document context so clause generation stays faithful.
- ▸LangSmith end to end—distributed tracing to benchmark performance (e.g. ~P50 0.87s, P99 < 5.2s sustained in our runs).
- ▸FastAPI async backend with streaming tokens, Google OAuth 2.0, custom rate limits, and a React front end. LangGraph reflection workflow for draft → critique → human-in-the-loop.
- LangGraph
- Supabase · pgvector
- FastAPI · React
- LangSmith · OAuth2
Munchy Mumbai
- ▸LangGraph semantic router classifies intent and dispatches to sub-agents to keep token spend disciplined.
- ▸Groq LPUs for fast inference; Supabase hybrid search (vector + trigram) to cross-check Llama 3 outputs and curb hallucinations.
- ▸Google Maps & YouTube APIs in a React SPA; FastAPI behind Docker, deployed to Vercel.
- LangGraph
- Groq · Supabase
- React · FastAPI
- Docker · Vercel
Amazon ML Challenge
- ▸Dual-model ensemble on 75K samples—fused Qwen-0.6B text embeddings with ViT vision features through a 3-layer regression head (~44.8% SMAPE).
- ▸Fine-tuned RoBERTa-Large FFN with a custom 3-layer head (GELU, Smooth L1); averaged both architectures for robust price prediction.
- ▸Competition placement: Rank 183 / 20,000+ teams (top ~1%).
- PyTorch · ViT
- Qwen · RoBERTa
- Multimodal · tabular
CF Time Tale
- ▸Estimator for total time on Codeforces—given a handle and derived profile signals (ratings, tags, contest activity, etc.), it predicts how much time you have sunk into the platform, not a toy single-problem timer.
- ▸LightGBM on engineered features, with heavy hyperparameter tuning; validation MAE 9.93 min and RMSE 14.06 min on the setup we reported.
- ▸Flask app behind nginx, Redis for caching/session-style speedups, containerised with Docker, deployed on Render with a dashboard for breakdowns and analysis.
- LightGBM
- Redis · nginx
- Flask · Docker
- Render · Python
More on GitHub
AURA: Cosmic Weather
Predictive analytics for space weather financial impact—FastAPI backend, React frontend, plus ML and Monte Carlo simulations to aid insurers.
React · FastAPI
ML Forecasting
Gemini AI
View code
ProSizeNet
Hybrid CNN-Transformer architecture achieving SOTA performance in predicting protein residue sizes; includes paths toward on-device inference with Core ML for Apple deployment.
CNN · Transformer
Core ML
Bioinformatics
View code
Multi-Agent Diagnosis
Multi-agent architecture for triage and routing: CrewAI + Gemini agents orchestrated behind a FastAPI service and a React UI.
Multi-agent
React · FastAPI
CrewAI · Gemini
View code
AI Job Assistant
Career co-pilot with smart job matching, ATS analysis, cover letters, and course suggestions—wired to Google (Gemini) APIs and complementary tooling for search and generation.
Streamlit · FastAPI
Google Gemini API
ATS · NLP
View code
SkillMorph
Career recommendation platform combining live job listings and course suggestions with auth—built with React, FastAPI, and Firebase for rapid hackathon iteration.
React · FastAPI
Firebase
Hackathon
View code
Paper Title Gen
Fine-tuned GPT-2 for research-style title generation using full fine-tuning and LoRA / QLoRA-style parameter-efficient training—highlights end-to-end model training, alongside the embedding, chunking, and RAG patterns I use in larger retrieval systems.
LoRA · QLoRA
GPT-2 fine-tuning
RAG · NLP
Student Heatmap
Interactive heatmap visualizing the distribution of Indian students using Pandas and Folium.
Shiny
Data Viz
Therapy Chatbot
Multi-agent mental wellness assistant with empathetic AI chat, journaling, and mood tracking.
Wellness
LLM
View code
Drug Target MLP
MLP model to classify drug target statuses using TF-IDF, achieving 95% validation accuracy.
Scikit-learn
TF-IDF
View code