Engineering LLM-powered systems — RAG pipelines, model distillation, and computer vision — and delivering them end to end, from FastAPI/Node backends to typed React frontends.
I build intelligent systems end to end — from the model to the interface. My core focus is Generative AI and machine learning: designing retrieval-augmented pipelines and LLM orchestration, compressing deep networks through knowledge distillation, and applying computer vision to real-world problems.
I pair that with full-stack engineering, turning models into deployed products backed by FastAPI or Node/Express services and typed React frontends. I care about substance over noise — models that are benchmarked, systems that ship to production, and code that reads cleanly for the next engineer.
Currently exploring LLM agents, efficient inference, vector databases, and the craft of taking research-grade ideas into reliable, usable software. Away from the keyboard, I'm usually on the cricket pitch or deep into a video game.
Languages
ML / AI
Frontend
Backend, Databases & Tools
An intelligent call-analysis platform that transcribes uploaded audio, generates AI summaries with sentiment, and lets users query their call history through a RAG chatbot. Deployed end-to-end with authentication, vector search, and a typed React frontend.
A RAG pipeline that verifies claims in AI/ML research papers — extracting assertions, retrieving supporting and contradicting evidence, and issuing transparent judgments to flag hallucinations. Benchmarked against the SkepticBench dataset.
A knowledge-distillation project that compresses a heavy image-enhancement model (MSAFN) into a lightweight student network (LightMSAFN) for real-time inference, achieving 51 dB PSNR and 0.9792 SSIM. Built on the teacher–student distillation work from my Intel Unnati training.
An AI-powered waste-classification system that detects and categorizes waste from uploaded images and returns disposal guidance, using a hybrid YOLOv8/YOLOv5 detection pipeline backed by a Flask API and a React interface.
A MERN-based invoice automation tool that uses Gemini-powered agents to extract invoice data from unstructured text and auto-generate payment-reminder emails, with JWT auth and full CRUD workflows.
I'm open to ML/AI and Software Engineering internship and new-grad opportunities. The fastest ways to reach me:
- Email: ayushsharma130408@gmail.com
- LinkedIn: ayush-sharma-6219352b1
- Portfolio: myportfolio-six-blond-97.vercel.app
- Kaggle: ayushs1308
