Chandrahas Aroori

• Architected multi model systems with Llama 8B for core tasks and Phi 2 2.7B for on-device file understanding. • Finetuned both models with knowledge distillation and QLoRA, achieving a 35% model size reduction and 70% inference speed improvement. • Online deployment of Llama 8B with vLLM on EC2 with elastic GPU • Improved model's natural styling with RLHF using Huggingface DPO.

• Designed and deployed Reinforcement Learning and Bayesian Optimization to optimize system resource allocation in ADAS (Qualcomm's Automated Driving Platform) , focusing on blackbox optimization techniques. • Built Low Power Sentry mode for ADAS (Automotive Driving Platform) that identifies threats to the system while consuming minimal power.

Developed and deployed multiple Machine Learning systems in the MMX Intelligence Team for various Microsoft Products (Microsoft ToDo, Planner & Project). • Engineered the “Picture to Tasks” feature, leveraging ResNet and BERT models to analyze images, and convert them into a detailed list of actionable tasks. Achieved 78% accuracy in picture to Task translation. • Lead the creation of “Discover Feed”, a dynamic smart feed that extracts actionable tasks from Emails and Teams messages using Hidden Markov Models. Contributed a 20% increase in user engagement with suggested tasks. • Lead developer for “Autosuggest Tasks” functionality, employing statistical models to analyze users' historical lists and provide personalized suggestions for new items. Contributed to a 25% boost in Task Creation. • Lead developer for “Autocategorize Tasks” feature, implementing NLP similarity models to automatically categorize various items enhancing user organization. Reduced manual categorization efforts by 35%.

• Improved HDMI I2C function callback in graphics driver which controls HDMI connection to GPU. • Enhanced the powercfg utility library to facilitate comprehensive testing of GPU power performance.

• Created a Mobile-First Platform with a scalable backend infrastructure, supporting up to 1 million users, featuring an Android app frontend and utilizing Spring Boot, Hibernate, and MySQL for the backend.

• Developed a drag & drop application platform for medical pathway providers, utilizing an Android app for the front-end and implementing the backend with Spring Boot, Hibernate, and MySQL.