Projects

Software, AI, ML, and analytics — each with a case-study page

Every project opens into a structured page: the problem, the architecture, the stack, the implementation decisions, the evidence, and demo links where they exist.

Looking for the automation products instead? Five systems are documented here →

01Featured

Strongest portfolio stories

Research / Full-Stack AI

Materia - Agentic Materials Simulation Platform

Embargoed

A FastAPI, React 19, and Electron platform where a 23-tool function-calling agent turns natural-language requests into real materials-science workflows: VASP/DFT input generation, structure tools, relaxation, MD, elastic, phonon, and NEB simulations.

~26K LOC across backend and frontend with Celery/Redis jobs, JWT auth, Fernet-encrypted key storage, Docker, and CI.
Validated 6 ML interatomic potentials across 60 relaxations at 2.65% mean volume deviation vs DFT references.
Gemini 2.5 Flash achieved 100% first-tool selection and argument accuracy on a curated 39-prompt suite.

FastAPI
React 19
Gemini
Celery
Redis
Electron
MACE
MatterSim

Case study →

Research / ML for Science

DKL-BO - Deep Kernel Bayesian Optimization

Embargoed

A CGCNN plus Gaussian-Process active-learning system for discovering high-band-gap 2D materials from a 3,351-crystal search space while evaluating only ~3% of candidates.

Found the global best material within ~3% of evaluations, with 6.3x efficiency over random search.
Validated across 30 seeds and a second 10K-crystal SNUMAT study with paired Wilcoxon p approx 0.002.
Built a ~9.2K LOC Hydra pipeline with LMDB graph caching, pool-embedding reuse, calibration metrics, and 5 test suites.

PyTorch
CGCNN
GPyTorch
BoTorch
Hydra
LMDB
Bayesian Optimization

Case study →

Software Engineering

Verdex - Full-Stack Online Judge

Complete & deployed

A LeetCode-style online judge where users solve C++ and Python problems in the browser, submit to a compile-once/run-all-cases judge, and receive AC, WA, TLE, RE, or CE verdicts with leaderboard tracking.

Full-stack judge with FastAPI, React/Vite, SQLite, Docker, GitHub Actions CI, and AWS EC2/ECR deployment.
Runs untrusted submissions in Docker sandboxes with networking off, 256 MB memory, 1 CPU, 64 PID caps, read-only FS, and rate limits.
Ships admin problem authoring, profiles, a leaderboard, Gemini AI code review, pagination, and 18 backend tests.

FastAPI
React
Vite
SQLite
Docker
AWS
GitHub Actions
Gemini

Case study →Watch demo Code

Data Science Competition

NFL Draft Prediction - Gradient-Boosting Ensemble

Deployed - Top 10% (GCI World 2026)

A LightGBM + CatBoost ensemble predicting whether a college athlete gets drafted into the NFL from Combine data, served as a live serverless API with SHAP explanations. Placed in the top 10% worldwide at GCI World 2026 (University of Tokyo, Matsuo Lab).

Reached 0.829 OOF ROC AUC with Optuna tuning (50 trials), 5-fold CV, 3-seed averaging, and an 85/15 LGBM-CatBoost blend.
Used leak-safe fold-internal target encoding for 236 schools and treated missing Combine drills as predictive signal.
Deployed as a serverless Modal /predict API with a Gradio UI, SHAP per-prediction factors, and Weights & Biases tracking.

LightGBM
CatBoost
Optuna
SHAP
Modal
Gradio
W&B
ROC AUC

Case study →Live API Code

GenAI / Agents

Sage - Agentic Multi-Tool RAG Chatbot

Complete & deployed

A deployed LangGraph assistant that decides per turn whether to answer directly or reach for a tool - web search, live stock quote, calculator, or semantic retrieval over your uploaded PDF - and remembers every conversation across restarts.

Built as a thin Streamlit client over a serverless Modal FastAPI backend (5 bearer-authed endpoints, scales to zero).
LangGraph conditional tool-execution graph routes across 4 tools with token streaming; Groq Llama-3.1-8B for inference.
Per-document RAG in Qdrant (thread-scoped payload filter), cross-restart memory via a Neon Postgres checkpointer, LangSmith tracing.

LangGraph
Groq
Qdrant
Neon
Modal
Streamlit
LangSmith
FastAPI

Case study →Live demo Code

02More work

Additional projects

Data Analytics

Retail Customer Segmentation & Intelligence

Analysis complete

A customer-intelligence project segmenting 3,900 retail shoppers by value, loyalty, and promotion dependency using Python feature engineering, SQL analysis, and Power BI reporting.

Engineered 10+ behavioral features and wrote ~15 SQL segmentation queries over a 3,900-customer enriched dataset.
Found that Champions - the top 25% value tier - drive ~49.7% of estimated annual revenue.
Delivered a Power BI dashboard, executive summary, and customer-retention playbook.

Python
SQL
Power BI
Pandas
Segmentation
Dashboarding

Case study →

Data Science / Analytics

Value-Weighted Telecom Churn Retention

Completed

An end-to-end churn-retention proposal that joins 100K telecom customers with usage records, predicts churn risk, and converts model scores into a value-weighted targeting strategy.

Selected XGBoost at ~0.70 ROC AUC after benchmarking Logistic Regression, Random Forest, and XGBoost.
Top risk decile showed 79.5% churn; targeting the top 20% captures ~30% of churners.
Projected $2.07M/yr net benefit and 2.1x ROI under stated retention-offer assumptions (modeled, not measured).

XGBoost
scikit-learn
Pandas
Matplotlib
Business Analytics

Case study →

GenAI / Coding Tools

Multi-Agent DSA Interview Assistant

Prototype

A Streamlit assistant where 6 specialized LLM agents collaborate to plan, brute-force, optimize, generate, review, and explain DSA interview solutions.

LangGraph pipeline with a reviewer-to-code-generator retry loop capped at 3 attempts over shared typed state.
Renders each reasoning stage in the UI so users can follow the problem-solving process.
Deployment, dependency pinning, and README polish remain planned.

LangGraph
Streamlit
Python
Agents

Case study →

GenAI / RAG

Indian Legal AI Assistant - RAG over Indian Law

In progress

A Streamlit RAG assistant that answers plain-English Indian legal questions from a corpus of legal PDFs and cites source documents and pages in the response.

Built a LangChain pipeline over 39 PDFs, 2,390 pages, and 9,851 chunks using bge-small embeddings.
Refactored the app into modular loader, embedding, vector-store, LLM, and config layers.
Remaining work: retrieval filtering and chat-history fixes, index persistence, and deployment.

LangChain
FAISS
Hugging Face
Streamlit
RAG
Python

Case study →

Recommendation Systems

Movie Recommender - NLP Content-Based Engine

Prototype

A content-based recommender that suggests 5 similar movies from a selected title using NLP tag vectors and cosine similarity over a 4,806-film dataset.

Precomputed a 4,806 x 4,806 similarity matrix with a simple Streamlit UI for title selection.
A quick-win project once it gets a reproducible notebook, posters, README, and deployment.
Intentionally listed as a prototype, not a flagship project.

Python
Pandas
Streamlit
Cosine Similarity
NLP

Case study →