Master's student, SCU | AI/ML | Building AI applications & RAG pipelines | Open to full time roles starting June 2026

github

medium

Khushi Savaliya  • Peerlist

• Developed a scalable Retrieval-Augmented Generation (RAG) pipeline to extract and retrieve insights from over 100+ GB PDF corpus, performing data extraction, semantic chunking, and embedding using Hugging Face’s BAAI/bge-base-en-v1.5 model.
• Implemented FAISS for vector search and evaluated IVFPQ indexing as an optimal approach (less than 100 ms latency) for large-scale retrieval, combined with query expansion techniques to improve recall.
• Deployed a zero-cost GenAI chatbot powered by DeepSeek LLM via OpenRouter, producing citation-backed, hallucination-resistant answers, and validated the pipeline’s economic and technical feasibility through performance benchmarking.

SEEK (Semantic Embedding & Extraction Kit)

• Built an end-to-end AI development assistant that autonomously plans, codes, tests, and documents software features from natural language requirements
• Architected a pipeline of 5 specialized agents using LangGraph for stateful workflow orchestration and integrated with GitHub for version control and PR generation
• Created a reinforcement learning feedback loop that improved code quality by 41% over time based on user reviews

Autonomous Software Development Agent

• Developed a question-answering system for efficient information retrieval, utilizing both abstractive and extractive approaches to generate and extract answers from Wikipedia’s history section
• For the abstractive model, employed ELI5 BART with the Pinecone vector database to perform semantic search and generate concise, human-like responses with a BLEU score of 0.85
• Implemented an extractive model using the Transformers library and a question-answering pipeline to identify precise answer spans within the text, reaching a 90% accuracy rate in answer extraction

Extractive and Abstractive QnA System

Applied ANN, SVM, and Decision Trees on a Kaggle heart disease dataset, achieving 94% accuracy through ensemble learning.

Heart Disease Prediction

• Enhanced lane detection and vehicle tracking accuracy by 92% by annotating and analyzing images and videos from the KITTI dataset using labelImg
• Developed and deployed YOLOv8 with PyTorch for real-time vehicle detection and curve tracking in lane-specific environments

Curve Tracking and Vehicle Detection in Laned

• Designed and implemented a hand gesture recognition system for controlling audio volume in real-time, utilizing OpenCV, Mediapipe for 21-point hand landmark detection, and Pycaw for integration with the Windows audio API.
• Enabled smooth volume adjustments by calculating the distance between thumb and index finger, reaching 95% recognition accuracy for responsive and intuitive human-computer interaction.