Things I built and deployed
Seven AI/ML systems — each with a live demo or public codebase, and a deployment story more interesting than the model training. Tap a card to view screenshots.
Production-grade resume analysis platform. Fine-tuned a Sentence Transformer on 9,700+ own synthetically generated resume-JD pairs, reducing MAE by 70%. Multi-layer NLP pipeline: Groq LLaMA 3.3-70B for structured extraction, spaCy NER for location privacy, RapidFuzz for keyword matching, live LinkedIn job ranking via Apify. Full-stack: FastAPI backend with dual JWT auth (HS256 + JWKS), Streamlit frontend, Supabase persistence, Docker, HuggingFace Spaces deployment.
Medical Q&A system grounded in source documents via retrieval-augmented generation. LangChain orchestration with Pinecone Serverless vector store (cosine similarity, k=3). Groq LLaMA 3.1 inference. Flask backend, Dockerized, deployed on AWS EC2 Ubuntu with automated CI/CD via GitHub Actions.
Arbitrary neural style transfer using Adaptive Instance Normalization. Frozen VGG19 encoder extracts multi-scale features (relu1–relu4). Custom AdaIN decoder trained from scratch on Google Colab T4 GPU. Combined content + multi-scale style loss. User-adjustable alpha blending. Deployed on HuggingFace Spaces via Flask + Gunicorn + Docker.
Real-time exercise form analysis and coaching using MediaPipe pose estimation with joint angle computation. FastAPI backend with WebSocket streaming for low-latency feedback. Groq LLM generates contextual coaching cues per rep. Tracks rep count, form quality score, and session metrics in real time.
Dual-biometric classroom attendance system combining face recognition (MTCNN + InceptionResnetV1 via facenet-pytorch) with voice verification (Resemblyzer speaker embeddings). SVC classifier with a single-student fallback path, Supabase backend for records, Streamlit frontend for live capture and review.
End-to-end NLP pipeline that pulls comments from any YouTube video via the Data API, cleans and preprocesses text, and classifies sentiment to surface audience reaction at a glance — with visual breakdowns of the positive / negative / neutral split. The AI Insight sentiment classifier is trained on own synthetically generated data, similar to the approach used in Criterion.
Conversational assistant that lets you upload a document and ask questions directly against its content. Built around a retrieval-based Q&A loop in Python, focused on quick local document grounding without a heavy infra footprint.