Shiven Paudyal

Real-time American Sign Language gesture recognition running entirely in the browser. Detects 29 ASL classes (A-Z plus SPACE, DELETE, NOTHING) using MediaPipe hand tracking and a custom-trained TensorFlow.js neural network. Achieves 95%+ frame-level accuracy with temporal smoothing across a 15-frame rolling window to handle ambiguous letter pairs (U/V, M/N, A/S). Built as the production implementation of my published B.Tech research on deep learning for sign language translation.

Full-stack autonomous agent that scrapes 1,200+ ML/DS job postings daily from Greenhouse, Lever, and LinkedIn, extracts structured data via Groq LLM, generates 384-dim embeddings for semantic resume matching, and emails a weekly market intelligence report. Includes a public job board, skill trend dashboard, and REST API.

Serverless AI career coaching agent built on Cloudflare's edge platform with persistent cross-session memory. Analyzes uploaded resumes to deliver structured ATS feedback, rewrites bullet points, and preps users for mock interviews — all powered by Llama 3.3 70B inference via Workers AI. Cross-session memory is backed by Durable Objects with SQLite, so context persists across browser refreshes with no traditional backend.

Real-time soccer player tracking system using YOLOv8 object detection with Kalman Filter for trajectory smoothing. Generates tactical heatmaps, speed metrics, and formation analysis from raw match footage.

Built on Roboflow's detection backbone — Kalman tracking, tactical heatmaps, and formation analysis are original implementations.

End-to-end ML pipeline predicting customer churn with 94% AUC. Features SHAP explainability dashboard for business stakeholders, MLflow experiment tracking, and automated retraining workflows.

Production-grade fraud detection system with full MLOps lifecycle — automated feature engineering, model versioning via MLflow, containerized REST API with Docker, and drift monitoring.