Mani Smaran Nair

Conducted advanced research in deep neural networks and variational inference to improve model generalization for regulatory region tasks at a leading cancer research center.

• Integrated a Foundational model using Borzoi-generated embeddings to train a deep neural network, significantly improving model generalization across diverse regulatory region tasks.
• Developed a modular guide class (AmortizedNormal) to seamlessly combine LLM embeddings with variational inference in NumPyro, streamlining complex model architectures.
• Performed scalable Stochastic Variational Inference (SVI) on over 20,000 DNA sequences, boosting ELBO convergence speed by ~30%.

Led the development and deployment of GenAI systems, focusing on real-time RAG pipelines and LLM-based solutions for knowledge extraction.

• Built and deployed a real-time Retrieval-Augmented Generation (RAG) pipeline with FastAPI, Sublime, and LangChain, retrieving enterprise knowledge from over 10,000 records.
• Collaborated cross-functionally on internal GenAI initiatives, including LLM-based solutions, to enhance knowledge extraction capabilities.
• Conducted user interviews, documented workflows, and facilitated workshops with cross-functional teams to gather requirements and ensure alignment for AI applications.
• Developed comprehensive test cases and automated evaluation pipelines for pilot AI applications, ensuring robust performance and reliability.
• Synthesized complex technical and business feedback into structured reports to guide strategic deployment and iterative improvements.

Analyzed patient-question interactions and enhanced predictive accuracy through advanced probabilistic modeling and data visualization techniques.

• Analyzed 15,000 patient-question interactions using advanced probabilistic response models to derive critical insights.
• Improved predictive accuracy by 18% compared to baseline cognitive models, demonstrating enhanced model performance.
• Utilized Seaborn and Matplotlib visualizations to conduct comprehensive model performance comparisons and present findings effectively.

Managed and automated fMRI data preprocessing pipelines for a large patient cohort, significantly enhancing efficiency and data consistency.

• Processed fMRI data from over 15,000 patients leveraging BIDS pipelines, ICA, and BET for precise brain extraction.
• Automated critical data preprocessing pipelines, reducing data preparation time by 50% and ensuring high consistency for research studies.

Developed a fullstack Retrieval-Augmented Generation (RAG) chatbot leveraging Langchain for intelligent document querying.

Implemented a food captioning model using a Diffusion-Augmented COCO Dataset to enhance image understanding and descriptive capabilities.