Data Scientist. Builder. Coder.

Fix with Me

Users type, talk or snap a photo of what is broken. A Bayesian diagnosis engine over a curated knowledge base picks the most informative next question, reaching 90.7% top-1 accuracy on a 94-case benchmark versus 81.3% for plain Gemini. Follow-up questions run through hybrid RAG on Supabase (pgvector + PGroonga, Reciprocal Rank Fusion, 100% recall@5) with per-claim verification by Gemma, cutting unsupported claims from 34% to 0%. A trained TF-IDF + logistic regression safety classifier (3.8K examples, 0.68 MB pure-Python runtime, about 0.2 ms per message) lifts hazard recall from 67.2% to 89.1%.

Web app on Cloudflare Workers, API on Render’s free tier: the first answer can take about a minute while the API wakes up.

MahaGuru AI

Two products that hand off to each other. StudentGPT is a reflective mentor that asks before it advises. It runs a deterministic English and Hinglish safety screen on every message, keeps a structured understanding record, and retrieves reference dialogues with BM25. Classroom turns a goal into a diagnostic, a versioned roadmap, generated lessons with an AI teacher, quizzes, rubric-graded projects and mastery tracking that adapts the plan. Built as a FastAPI and React monorepo with streamed replies, guest-to-account sessions and a provider-agnostic LLM layer (Gemini by default). Shipped as one Docker image with CI on every push.

Hosted on Render’s free tier. The first load can take about a minute to wake up, and AI replies use a daily free quota.

Customer Churn Prediction

XGBoost and LightGBM trained on ~594K telecom records (21 features) with 5-fold stratified cross-validation; XGBoost reached a 0.916 mean ROC-AUC. EDA surfaced the strongest churn drivers — electronic-check payment (49%), month-to-month contracts (42%), fiber-optic internet (41.5%) and first-year customers (49%). Deployed as a live app with batch scoring and per-customer explanations.

Hosted on Render’s free tier — the first load can take ~30 s to wake up.

CreditWise — Loan Approval

A scikit-learn pipeline (imputation, scaling, one-hot encoding) on 950 applicant records with 18 features. A class-balanced Random Forest reached 0.91 F1 and 0.98 ROC-AUC on a stratified 20% test set, versus 0.81 F1 for a logistic-regression baseline. The Streamlit app returns live predictions with confidence scores and feature importance — credit score and DTI ratio are the strongest predictors.

IAMARS — Drone Detection & Tracking

Led a team that fine-tuned YOLOv8n on the VisioDECT drone dataset with a leakage-aware block split, reaching 0.965 mAP@0.5 on a 1,800-image held-out test set. The pipeline runs at 88.5 FPS on an RTX 2050 and 20 FPS on CPU; tracking was evaluated with MOTA/IDF1 on 91 test videos, plus intercept-point prediction in 3D simulation. Found and fixed a label-conversion bug that had inflated training mAP.