Ashwin Iyer

Designed a highly effective Transformer architecture based Recommendation system using BERT, Databricks, MLFlow, Azure Delta Lakes and Azure DevOps that processes 10 million recommendations in under 1 minute with no pre-training requirements. Conceptualized an extremely performant Video Understanding Engine, that uses Multimodal and Generative AI, in-house LLMs, Docker and FastAPI, to extract plot, summary, cast and various other metadata from just the video, faster than realtime (approx. 0.3-0.5x duration of the video). Implemented a RAG using Knowledge Graphs, LLMs (Mistral 7B), LangChain and FastAPI that can be used to perform natural language search instantaneously. Developed a powerful Generative AI system to automate workflows with SOPs and technical requirements, employing Multi-Agent Frameworks; achieved 40% efficiency boost in routine operations.

Built a privacy-preserving Federated Learning environment to classify Alzheimer’s using CNNs and Siamese Networks with 46% accuracy. Implemented custom optimization algorithms such as FedProx, FedNova and FedAvg for federated training with an average triplet loss of 1.2 per round.