AI / ML Researcher
Electronics & Telecommunications Engineer working at the intersection of large language model efficiency and edge deployment. Research focus: quantization, inference-time compute, and post-training optimization.
I'm a researcher and engineer with a background in Electronics & Telecommunications Engineering, specialising in making large language models smaller, faster, and deployable on constrained hardware. My work sits at the intersection of cs.LG and cs.AR.
I have production experience building RAG systems and quantized LLM pipelines, a preprint on LLM quantization, and a Kaggle competition placement. I'm currently pursuing MSc AI/ML programs abroad with a focus on efficient inference research.
Longer term, I'm interested in founding AI-native tools and contributing to open-source infrastructure for efficient ML.
arXiv preprint
Research on scalable methods for editing knowledge in 4-bit quantized large language models, addressing the precision-accuracy tradeoffs in efficient LLM deployment.
In Progress
Benchmarking Llama-3.2 and SmolLM2 on ARM architectures to evaluate the viability of 1-bit quantization for mobile edge deployment, exploring the efficiency-accuracy tradeoff.
IJRASET, DOI: 10.22214/ijraset.2024.64932
Analyzing the correlation between meteorological and geographical factors and flood severity in Myanmar using machine learning approaches.
A selection of research prototypes, open-source tools, and engineering projects.
Achieved a top 17% global ranking (194/1142) by optimizing Python/C++ code for extreme brevity and algorithmic efficiency. Solved complex algorithmic challenges under strict character-count constraints, demonstrating mastery of low-level optimization.
Developed a text classification model to detect toxic content on Quora, handling a dataset of 1.3M+ questions. Implemented GloVe embeddings and LSTM networks, achieving an F1-score of 0.68 on a highly imbalanced dataset.
Architected a Retrieval-Augmented Generation pipeline for HR automation at Asanify Technologies. Reduced query response latency by 35% and automated 80% of Level-1 employee support tickets using quantized LLMs.
Built and deployed predictive models processing 50,000+ records at Source Code Technologies, improving data classification accuracy by 15%. Containerized ML models using Docker for cloud deployment, ensuring 99.9% system uptime.
Asanify Technologies, Pune, India
Source Code Technologies Pvt Ltd, Pune, India
Advanced Software Engineering Job Simulation
Jan 2024 – June 2024Software Engineering Job Simulation
June 2024Intro to AI Ethics
July 2024 – Aug 2024Rating: 1568 (Top 28.89% worldwide)
Ranked 6976th among 20k+ participants
Ranked 194 out of 1142 participants (Top 17%)
I'm open to research collaborations, PhD discussions, industry roles in AI/ML, and conversations about efficient inference. Reach out via email or any of the links below.
kabirpotdar7@gmail.com