Selected work
F1 Telemetry Lens
Driver fingerprinting from raw telemetry — identifying F1 drivers by how they brake, corner, and throttle. Live Streamlit dashboard.
The problem
Can you identify a Formula 1 driver purely from telemetry — speed, throttle, brake, gear — with no idea who was in the car? I built a system to find out.
What I built
End-to-end pipeline using FastF1 for lap telemetry, with 9 engineered driving-style features per lap (braking signatures, corner speed, coasting ratio). XGBoost classifier at 93.7% 5-fold OOF accuracy; SHAP revealed mean_corner_speed and coasting_ratio as the most discriminative. A 1D-CNN encoder (40k params) produces 32-dim lap embeddings — UMAP shows perfectly separated per-driver clusters at silhouette 0.84. Deployed as an interactive Streamlit app with radar profiles, UMAP explorer, and a live blind identification challenge.
Result
Nasdaq Stock Predictor
End-to-end MLOps pipeline — real-time ingestion, XGBoost forecasting, drift detection, Prefect orchestration, and a Streamlit monitoring dashboard.
The problem
Most ML projects stop at a trained model. This one is built like production: scheduled ingestion, data validation, observable pipeline runs, and a containerised deployment you can actually hand to someone else.
What I built
Real-time NASDAQ data ingestion via Yahoo Finance API into PostgreSQL with idempotent transformations. Great Expectations for data quality validation. XGBoost forecasting model integrated into the pipeline. Prefect for scheduled orchestration with retries and observable run history. Full Docker containerisation. Streamlit dashboard with structured logging for monitoring and error tracking.
Result
Sign Language Translation System
Fine-tuned T5 for text-to-ASL-gloss translation with video synthesis — supporting 4 input modalities. Deployed as a PWA with Firebase auth.
The problem
Deaf and hard-of-hearing users face a communication gap that text alone doesn't solve. This was my FAST-NUCES FYP — a system that translates any input (text, speech, image, or video) into ASL sign video.
What I built
Fine-tuned a T5 transformer for text-to-ASL-gloss translation, then built an FFmpeg video synthesis pipeline to map generated gloss sequences to sign video frames. Four input modalities supported: typed text, audio (via AssemblyAI), image (Google Cloud Vision OCR), and YouTube video URLs. Deployed as a Progressive Web App with Firebase authentication and a React frontend.
Result
Ride Intelligence API
Production-style REST API in Java + Spring Boot with a rule-based fraud detection engine scoring rides across 4 anomaly patterns.
The problem
A REST API built to demonstrate production-grade Java design — layered architecture, anomaly detection, and proper test coverage, not just working endpoints.
What I built
REST API in Java 8 and Spring Boot following Controller → Service → Repository layered architecture. A rule-based anomaly detection engine scores each ride across 4 fraud patterns and returns a composite risk score. Unit and integration tests written with JUnit 5 and MockMvc — covering both happy paths and edge cases.
Result
MedTrack
Mobile app for medication schedule management — reminders, refill tracking, community blogs, and analytics. Android + Firebase with a React admin panel.
The problem
Medication non-adherence is a real health problem. MedTrack makes it frictionless — smart scheduling with four frequency modes, refill threshold alerts before you run out, and adherence analytics over time.
What I built
Android app (Java) with Firebase Authentication, Firestore database, and FCM push notifications for medication and refill reminders. Medication scheduling supports four modes: once daily, twice daily, interval (every N hours), and specific days with date ranges. Community tab with blog publishing (rich-text editor, admin approval flow), ratings, and reviews. Adherence analytics on the profile tab. Separate React admin panel for blog moderation and reporting.
Result
Marula & Sage
Full e-commerce storefront for a fictional clothing brand — shop, cart, checkout, wishlist, search, and reviews. React frontend with a hand-built design system.
The problem
Most frontend demos show a product grid and call it done. I wanted to build the full shape of a real storefront — every flow a customer or store owner would actually expect, not just the parts that are easy to show off.
What I built
A 22-product catalog across 5 categories with a custom design system (type, color, and a hand-drawn SVG signature motif rather than default UI). Full cart and checkout flow with client-side validation, order confirmation, and shipping logic. Search, category and price sorting, a wishlist, star ratings with a review breakdown, and an interactive size guide. Built in React with Vite and React Router, no backend — state lives in context, demo orders persist to sessionStorage.
Result