Computational Neuroscience & Artificial Intelligence
MS by Research, IIT Madras · advised by V. Srinivasa Chakravarthy
Previously AI Engineer & Researcher at Neurogati
I work on representations that carry structure — dynamics, prediction, geometry — rather than capacity alone. Right now that mostly means oscillators: models of neural computation where the dynamics, not the weights alone, carry the representation, tested against EEG and fMRI responses to naturalistic stimuli. The applied side is clinical — Parkinson's biomarkers from video, audio, and gait, and markerless 3D pose running in a physiotherapy clinic.
Currently
experience & education
IIT Madras
July 2026 · Present
Research in the Computational Neuroscience Lab (Chakravarthy Lab), on oscillatory and predictive models of neural computation.
Training on banks of Hopf oscillators, and testing them against EEG and fMRI responses to naturalistic stimuli.
Self-supervised objectives for spatial memory in grid-world navigation work.
Neurogati (Incubated at IIT Madras Research Park)
Jun 2025 — Jun 2026
Founding engineer on the AI team, building clinical measurement systems from research prototype through to deployment in partner clinics.
QuamonPD — a digital biomarker platform for Parkinson's disease that quantifies motor and speech symptoms from phone video and audio. Built the vision and audio pipelines and the local-first mobile capture architecture for use in clinical settings with unreliable connectivity.
StereoPose3D — markerless 3D motion capture for physiotherapy, recovering metric joint positions from a calibrated stereo camera pair. Deployed and in ongoing use at a partner clinic.
Spatial navigation and memory — extended the grid-world graph work into a scene-graph memory system, pairing a Neo4j graph store with vision-language models so an agent could answer questions about an environment it had explored.
IIT Madras
May — Jul 2024
Developed a system that uses Graph Database, Graph Neural Networks and Large Language Models(LLMs) to navigate a grid world, answer user queries, and provide relevant information about the environment.
Processed the dataset to store it in a Neo4j graph database and for other downstream tasks.
Used the LLM(Llama3) to question the graph database and generate Cypher queries for the questions asked by the user. The prompts were dynamically tuned using Langchain to improve query generation.
Implemented Q-learning method to explore the grid world. Used the explored data to query the graph neural network using the LLM(Llama3).
Presented this work at the 5th Annual Research Showcase, IIT Madras.
Bungee Tech
Jun 2023 — Jul 2023
Matched retail pet and grocery products to market conditions for effective pricing.
Worked across a 5TB dataset of over 100 million products, with attributes spanning titles, images, descriptions, sizes, flavours, model numbers, and UPCs.
Ran exploratory analysis on text and string data using Python libraries and SQL queries via AWS Athena.
Built a Transformer-based architecture with approximate KNN, benchmarked at 96% F1 on the WDC product matching benchmark and 91% on an internal evaluation set.
SASTRA Deemed University
2021 — 2025
Core coursework in Data Structures, Algorithms, and Machine Learning.
selected work
Research
Physics-Grounded Deep Learning & Signal Processing
Physics-grounded autoencoders that use Hopf oscillator dynamics across EEG, audio, and fMRI, grounding representation learning in dynamical-systems theory.
Multi-Agent AI & Cognitive Systems
A multi-agent cognitive assistant built on a dual-memory architecture, combining short- and long-term memory with spatial reasoning to plan and act across complex tasks.
Generative AI & Music
A code replication of Google's MusicLM, trained on an A100 GPU, using three custom transformer models that sequentially convert a text prompt or audio input into a musical output.
Shipped
AI in Healthcare & Digital Biomarkers
Quantifies motor and speech symptoms from video and audio to support clinical assessment and remote monitoring. Currently moving toward clinical pilot testing.
Computer Vision & 3D
A stereo-vision pipeline that recovers 3D pose from synchronized camera views — detecting 2D keypoints on each view of a calibrated stereo pair and triangulating them into metric 3D positions.
3D Reconstruction & Computer Vision
An end-to-end pipeline for multi-view 3D reconstruction with super-resolution, built for a Kaggle competition — pairing feed-forward reconstruction with detail enhancement and Gaussian-splat rendering.
Papers & Talks
Deep Oscillatory Neural Networks for EEG signal compression
Computational Neuroscience Lab, IIT Madras · 2026.
Oscillatory representation learning for fMRI encoding of naturalistic stimuli
Computational Neuroscience Lab, IIT Madras · 2026.
Is deep learning enough for handling X-ray images? A benchmarking study
SASTRA
Interactive Captioning of Medical Images using Large Language Multimodal Models
SASTRA
An Efficient LLM for the Indian Judicial System – IndicLegal-Mistral
SASTRA
Graph databases and large language models for grid-world spatial navigation
5th Annual Research Showcase, IIT Madras, 2024
Tools & Domains
Outside the Lab
Mostly available light: architecture, long exposures, cities at hours when they aren't performing. The discipline is subtraction — deciding what stays out of frame.
Blender and Unreal. Modelling and lighting a scene is the one place I think spatially instead of symbolically, which turns out to transfer more often than expected.
Pencil, ink, tablet. Mostly figures and linework. It's the slowest feedback loop I have, and the only one I can't debug my way out of.
Classically trained pianist, Trinity Grade 8, working in orchestration and Carnatic music. Most of what I understand about frequency, resonance, and phase I met here before I met it in a paper.
get in touch
Copyright Shrinivas Sesadri 2026
Created by Shrinivas Sesadri