Tulika Awalgaonkar

About Tulika Awalgaonkar

I am interested in discovering and proposing the fundamental principles, algorithms, and implementations for solving software development-related problems. Over the past years, I have worked on several projects in the Android, Web, and machine learning domains. I worked as a Software Engineering Intern at VMware in their vSphere Bitfusion team. I was working on analyzing Bitfusion performance for different ML benchmark usage scenarios, architecture, and AI frameworks. I worked as a founding engineer at Lamini, which is working on building specialized LLMs for enterprises. I worked on moving their current system to serverless using Azure functions and building fine-tuning pipelines using Slurm. I have also worked on the optimization of training and inference of LLM models using techniques like LORA, Deepspeed, and Accelerate. Currently, I am working as a Machine Learning Engineer at Salesforce.

Experience

1. Machine Learning Software Engineer - Ai Research

   Salesforce

   Jan 2024 — Present · Palo Alto, CA, US

   Developed xLAMFlow in Swift, an iOS productivity assistant powered by quantized xLAM-2, supporting single-step, multi-turn, and multi-step agentic scenarios demonstrating on-device AI, privacy-preserving inference and end-to-end user experience integration. • Advanced model training pipelines by preprocessing datasets (AgentBank, ToolACE, BFCL) and fine-tuning Qwen-1B/3B/8B and Llama3-8B, released as xLAM-2, then quantized to GGUF format for edge deployment.• Developed a feedback-driven pipeline with external annotators to assess and refine agentic data, updating 25.6% of samples for improved fine-tuning robustness.• Developed a feedback-driven pipeline with external annotators to assess and refine agentic data, updating 25.6% of samples for improved fine-tuning robustness.• Designed and open-sourced MobileAIBench, a cross-platform iOS/Android benchmarking framework to evaluate on-device LLM/LMM performance, leveraging llama.cpp quantization (4-bit, 8-bit, 16-bit) for efficiency.• Optimized multimodal inference by quantizing and deploying internal xGen-MM models on-device, enabling low-latency multimodal reasoning directly on mobile hardware.

Education

• UC Irvine Donald Bren School of Information and Computer Sciences

  Master of Science - MS, Computer Science

• Pune Institute of Computer Technology

  Bachelor of Engineering - BE, Computer Engineering

  2016 — 2020