TensorFlow, PyTorch, JAX, Scikit-learn, Pandas, Numpy
Fine-tuning open source models such as google flan t5 series on different datasets and then deploying them to production.
I have worked with opencv, fine tuning vgg16 for real time detection, working with both convolutional neural network and vision transformers.
FARM stack, Django, Next.js, React, Flask, FastAPI, building scalable web applications and APIs mostly for deploying LLMs or other ml models in production
I have used Azure, AWS, railway, koyeb services for deploying ML models, serverless computing, and cloud infrastructure.
Integration with OpenAI, Gemini, Claude, and Hugging Face APIs for advanced AI applications.
Plotly, Matplotlib for data visualization; SQL and NoSQL databases like PostgreSQL and MongoDB.
A monolithic progressive web app in Django with PostgreSQL as database and a complete agentic AI content generator using LangGraph, scheduling and notifying users before posting via cronjobs and resend API, with social authentication options like Google sign in, cloudinary api for users media, and security with cloudflare.
Developed with FARM stack, role-based access control, Gemini API, LangGraph, Cloudinary for images, MongoAtlas DB, WhatsApp and email notifications.
Created using Next.js and MongoDB, property listings and blogs, user submissions with admin approval, CRUD operations for properties and blogs, separate admin panel.
A dynamic and modern Next.js built website for construction with WhatsApp API connected and MongoDB database.
Used deep learning architectures to detect commercial starts and categorize ads. Detailed pipeline available, source code under NDA.
Used DBSCAN to find relative average predictions compared to bookmakers, detects errors if variance exceeds threshold.
Chess engine using encoder-only transformer, UCB, and MCTS in ensemble for real-time play via terminal.
Fine-tuned Google BERT for analyzing sentiment in user comments and responses.
Fine-tuned BERT with grammar and spelling checks to evaluate essays with IELTS band scores.
Fine-tuned Llama model on custom dataset to answer physics-related questions accurately.
Decoder-only architecture for predicting and generating language sequences.
Implemented Vision Transformer with VQ-GAN for image generation, compared with StyleGAN and diffusion models.