Data Science Intern at ADNOC- Abu Dhabi, UAE (Jan ‘25 - Mar ‘25)
Automated forecasting, cutting manual research time by 90%, by developing a pipeline to fetch, analyze, and score emerging technologies.
Implemented an LLM-powered evaluation system, to summarize and rank innovations, delivering real-time insights via a dashboard.
Machine Learning Intern at Proshort- Bangalore, India (Jun ‘24 - Jul ‘24)
Fine-tuning and quantization of advanced large language models (LLaMA-3-8B and Phi3-mini-128k-Instruct) using state-of-the-art tools including Hugging Face, unsloth and Weights & Biases, achieving significant improvements in model efficiency and performance.
Deployed quantized models through the implementation of Docker, ollama and FastAPI, ensuring robust, scalable, and efficient model deployment and usability.
Conducted comprehensive evaluations using ROUGE-2 score and also by Retrieval-Augmented Generation (RAG) method, integrating RAGA’s evaluation techniques to rigorously assess and enhance model quality and accuracy.
Designed and optimized the LangGraph pipeline, alongside sophisticated user session management workflows, to streamline complex language processing tasks and enhance overall user experience and interaction continuity.
Intern at Denso International India- Gurugram, India (May ‘23 - Jul ‘23)
Engineered a robust software solution for analyzing vehicle malfunctions using ECU data for Suzuki Motors. The software accurately displayed error codes and provided dynamic data visualization for enhanced diagnostics.
Developed a software application utilizing Pytesseract for accurate character recognition in engineering model drawings, streamlining the process of digitizing and interpreting technical information.
Created a comprehensive software application for Advanced Driver-Assistance Systems (ADAS) data collection. Integrated camera data with GPS information to facilitate precise data viewing and synchronization.
Research Intern at Santa Fe Research- IIT Madras, India (Jun ‘22-Aug ‘22)
Conducted a research project focused on optimizing vehicle routing with constraints on time, capacity, and dimensions. Developed a feasible solution using Google’s OR-Tools and seamlessly integrated 3D-bin packing for item arrangement within vehicle containers using the Py3dbp library.
Improving the quantitative reasoning capabilities of text-to-image models, specifically addressing challenges in counting accuracy, size proportionality, and fractional representation.
Curated a benchmark dataset of 350 high-quality prompt-image pairs.
Fine-tuned Stable Diffusion XL using LoRA adaptations and Direct Preference Optimization (DPO), achieving substantial improvements in metrics.