Leading a team of 6 awesome engineers to improve AI-tutor.

Product owner for AI-tutor from its inception to initial release - https://schoolai.coschool.ai/Built AI-tutor around three key pillars1) Accuracy (Conversational and Numerical) - Developed a Mixture of Experts model, achieving ~10% improvement over GPT-4 benchmarks.- Employed A/B testing to tune model parameters for conversational accuracy.2) Guard-railing- Used Retrieval Augmented Generation (RAG) to ground LLM knowledge.- Implemented computer vision techniques to extract data from non-parsable PDFs. - Ingested the data (text, images and videos) in vectorDBs using semantic chunking.3) Cost Optimisation- Reduced dependency on GPT by training BERT models using GPT logs.- Fine-tuned open-source LLMs (Mistral, Llama) using AWS Sagemaker and Hugging Face.

1) Set up an end-to-end voice service that generated, cached and played audio on demand. Enabled live captioning with millisecond accuracy. Reduced cost by $60k by implementing an asynchronous model.2) Created a ML model for semantic question duplicity check. Zero human interference required; automatic threshold calculation and para-phrase generators ensured a robust and scalable solution. 10x faster and 100x more cost-effective than GPT3

Working under Prof. Kavita Vemuri to investigate region specific activation of the brain during sustained deep breathing using a custom EEG dataset

Teaching assistant - Spring '22 - HS8.102 Introduction to Human Science

Working as an undergraduate research student at the Cognitive Science Lab with Prof. S Bapi Raju to explore methods for characterising the complexity of brain dynamics using EEG and fMRI signals.