Umesh Ghaskata

AI Shopping Agent

Personal Project | #FastAPI #OpenAI #SerpAPI #Pydantic #AWS #Uvicorn

Built an AI-powered shopping assistant that accepts natural language product queries and a budget, searches Google Shopping via SerpAPI, filters results by price, and uses an OpenAI model to compare the shortlisted products and recommend the best option with reasoning. The frontend is served by FastAPI and the backend is designed for public deployment on AWS EC2.

Tech Stack: Python, FastAPI, Uvicorn, OpenAI GPT-4o Mini, SerpAPI, Pydantic, HTTPX, Requests, python-dotenv, AWS EC2.

Multimodal RAG Platform

Personal Project | #Python #FastAPI #OpenAI #RAG #Supabase #pgvector #Docker

Built a production-ready Multimodal RAG platform that answers questions from PDFs, DOCX files, and images using semantic search and LLMs. Developed an end-to-end ingestion pipeline with OCR, image extraction, AI captioning, intelligent chunking, embedding generation, and vector indexing. Implemented context-aware retrieval using OpenAI Embeddings and Supabase pgvector with cosine similarity search, enabling accurate responses with document citations.

Supervised Fine-Tuning Gemma 3-1B LLM with LoRA vs QLoRA

Personal Project | #LLM #LoRA #QLoRA #AWS #PyTorch #HuggingFace

Experimented with parameter-efficient fine-tuning (PEFT) by training Gemma 3 1B IT on a custom Smoking Cessation / Motivational Coaching dataset (~1,500 conversations) on AWS EC2 (Tesla T4). Compared LoRA and QLoRA (4-bit NF4 quantization) trade-offs using BERTScore and ROUGE-L metrics. Both methods significantly outperformed the base model.

GenAI-Powered Career Readiness Platform

Personal Project | #GenerativeAI #React #Node.js #MongoDB #GoogleGemini

Built a career preparation platform that uses Generative AI to analyze resumes against job descriptions, identify skill gaps, and generate personalized interview preparation plans. Leveraged Google Gemini to create ATS-friendly resume content, technical/behavioral interview questions, candidate-job match scores, and structured learning roadmaps.

Tech Stack: React.js, Node.js, Express.js, MongoDB, Google Gemini, JWT, Zod, Axios, Puppeteer, PDF-Parse.

Image-Caption-Generator

University at Buffalo | #DeepLearning #CNN #LSTM #ComputerVision #NLP

End-to-end Deep Learning-based Image Caption Generator that automatically generates natural language descriptions for images. Combines Computer Vision and NLP using a pre-trained CNN for feature extraction and an LSTM-based sequence model for caption generation, learning to produce meaningful captions word-by-word from input images.

Autonomous Drone for Search Operations

IIT Bombay | #Drone #ROS #OpenCV

Developed an autonomous drone using Pixhawk + RPi, HOG-based detection, and A* algorithm with ultrasonic obstacle avoidance.

ML Projects: Logistic Regression, CNNs & RL

SUNY Buffalo | #PyTorch #ReinforcementLearning #CNN

Implemented logistic regression from scratch using gradient descent and L2 regularization. Built and tuned CNNs and MLPs in PyTorch with dropout, batch normalization, LR scheduling. Designed a custom RL environment and trained agents using SARSA and Double Q-learning.

Bollywood Celebrity Recognition

IIT Bombay | #TensorFlow #MTCNN #Streamlit

Built a facial recognition app using VGGFace (ResNet50) and MTCNN for face detection and embedding. Stored 2048-feature vectors using Pickle. Deployed a web app using Streamlit for real-time celebrity classification.

Coordinated UAVs for Efficient Spraying

IIT Bombay | #AStar #TSP #PathPlanning #OpenCV

Modeled a 2D grid-based spraying path using OpenCV and HSV heatmaps. Applied Traveling Salesman Problem to optimize waypoint sequence and simulated A* and Dijkstra algorithms for flight efficiency.

Operating Systems Programming

IIT Bombay | #xv6 #Shell #MemoryManagement

Built a Linux shell using fork(), exec(), wait() and implemented dynamic memory management in xv6. Used pthreads, mutexes, and semaphores for multi-threaded synchronization in user-space programs.