Designed and built the official marketing website for Dhanlakshmi Biochem end-to-end — a clean, responsive React + Vite site that establishes the company's online presence and showcases their product range to B2B customers.

• Modern, responsive UI built with React + Vite and modular components
• Dedicated product / category sections with a content layout the team can update without touching code
• Enquiry / contact flow wired to the client's preferred channel
• Optimised for performance, accessibility and SEO so the site loads fast and ranks well

AI-powered networking platform that finds the best people to collaborate with on GitHub and LinkedIn, given your resume, technical skills, and networking goals. Upload a resume, answer five quick questions, and the system returns ranked, personalised matches with reasons.

• Resume OCR with EasyOCR + PyMuPDF to extract skills and experience from PDF / image resumes
• Keyword extraction and match scoring using Ollama locally and Groq's Llama 3.1 70B in the cloud
• GitHub matcher analyses repos, tech stack and activity; LinkedIn search powered by a Playwright-based scraper
• 6-service architecture (Next.js frontend, FastAPI services, Ollama) deployed as a single docker compose stack

End-to-end deep-learning pipeline that classifies Arabic speech as native vs non-native using Wav2Vec2 prosody embeddings. Built for Hackenza 2026 and deployed live on Hugging Face Spaces.

• Uses facebook/wav2vec2-lv-60-espeak-cv-ft and concatenates mid-layer (4 / 6 / 8 / 10) temporal statistics with pitch features for prosody-focused 2051-d embeddings
• Three model heads — MLP baseline, MLP with SE-Attention, and a LightGBM gradient-boosted ensemble
• Validation-based threshold sweep to handle class imbalance (70 / 30 split); best ensemble hits ~85% accuracy / 0.82 F1
• Automated pipeline: download → chunk → embed → train → evaluate, with clean script-by-script stages

A Chrome (MV3) extension + FastAPI backend that turns any PDF you have open in a browser tab into a Report, Presentation, Podcast, Quiz, or interactive chat — all powered by NotebookLM, with per-user authentication that requires zero server-side credential management.

• Per-user auth model: the extension reads each user's Google cookies via chrome.cookies and forwards them as an X-NLM-Auth header — no shared session token on the server
• Backend creates a per-user NotebookLM client cached for one hour, so the developer deploys once and never has to refresh session tokens
• Generates Report (PDF), Presentation (PPTX), Podcast (MP4), Quiz (PDF) and supports chat with the PDF + Google Docs export
• FastAPI service containerised with Docker; deploy guide for Render included in the repo