Nishan Srishankar

I am currently a Senior AI Researcher at the AI Research Lab at JPMorgan Chase.

I was previously a Data Scientist/ Machine Learning Engineer 5 at Fidelity AI and worked on applied research for document understanding, and speech recognition.

I interned at the Honda Research Institute working on explainable AI, at the SETI Frontier Development Lab on self-supervision/knowledge discovery.
I worked on time-series understanding and action prediction with Warfighter Analytics for Smartphone Healthcare and full-stack DL applied to sustainability with the Army Research Laboratory.

I graduated with a MSc. in Robotics at Worcester Polytechnic Institute working with Prof. Carlo Pinciroli at the Novel Engineering for Swarm Technologies (NEST) Lab. My research involved developing decentralized algorithms for collective spatial perception in a robotic swarm given adversarial conditions and agents.

I also graduated with a Bachelors in Mechanical Engineering with minors in Aerospace and Electrical Engineering. My undergraduate thesis was to prototype a flapping-wing Micro-Aerial Vehicle.

Email  /  Resume  /  Scholar  /  Github

Research

I am interested in the intersection of robotics, machine learning, and human-centered AI. I am particularly interested in developing algorithms for multi-agent systems, decentralized control, and reinforcement learning. More recently, I have been tying together my work on swarm robotics with LLMs through building systems with robust, heterogeneous, long-term LLM agents. I am also trying to break into the (representation) alignment and robustness/trustworthiness fields.

We build an agentic framework and system that uses domain-specific APIs to extract and analyze information asked by users regarding contracts.

We preempt the chance that an LLM can misunderstand a query and yield incorrect outputs by using a model trained using Monte-Carlo simulations and a query rewriting process.

We develop a framework for easy and successful adapting of GUI/web agents to unseen tasks using 1-2 demonstrations in similar domains.

We construct a heterogeneous multi-agent swarm that answers a query about financial documents by validating the question, breaking it into sub-tasks, and ochestrating the sub-task to agents specialized on particular filings.

We propose a model that builds human intepretable explanations within the context of multi-agent driving by inferring the underlying system dynamics and agents' relationships using an interaction graph.

We utilize ray-traced rendering of Blender to reduce the sim2real gap and synthesize documents in randomized natural scenes for training downstream models.

We develop a framework that generates an interpretable object-level interaction graph by grounding the low-dimensional latent space representation on semantic behaviors with expert domain knowledge and apply this to highway driving behavior.

We develop data-driven decentralized federated learning for swarm robots without a central aggregating server through a gossip-based conflict-free replicated data structure (CRDT).

We use self-supervision on unlabeled multispectral Earth Observation data to develop a semantic representation encoder for knowledge discovery.

We explore AI-driven approaches and the need for open-source development to reduce manual labor and aid future lunar/ planetary exploration through real-time data analysis.

We present a novel parametric method for real-time, realistic, multi-manual grasping and interaction with complex objects.

We propose a framework which allows simulation of any generic, user-specified soft-body and interaction with a variety of input interface devices.

Creation of a transonic wind tunnel to to work between M=0.3 to M=0.9 for downstream testing.

Miscellanea

1. Method and system for improving code generation quality of Large Language Models through code guardrails. Patent filed.
2. Method and system for information extraction and aggregation. Patent filed.
3. Method and system of training an encoder classifier model in predicting hallucination of a machine learning (ML) model before a generation of a query. Patent filed.
4. Method and system for adapting web agents to new tasks using few human demonstrations. Patent filed.

1. ICRA 2016 Formal Methods in Robotics Scaling Chain of Integrators Winning team
2. WPI International Scholarship & Dean's List
3. Charles O. Thompson Scholar
4. Tau Beta Pi (Engineering Honor Society)
5. Edexcel Challenge Trophy & Best Academic Results 2011

1. Reviewer/ Program Committee
   1. CDC17
   2. AAMAS {19, 20}
   3. NeurIPS {20, 21, 22, 23}
   4. IEEE-RAL {20, 21, 22}
   5. ICLR {21, 22}
   6. AAAI {21, 24}
   7. ACML22
   8. IROS {22, 23}
   9. ICRA23
   10. ICAIF23
   11. AABI24
2. Other
   1. Panel discussion member, 2024

1. Artificial Intelligence, Fall 2018, Professor Dmitry Korkin
2. Introduction to Communication & Networks, Spring 2018, Professor Alexander Wyglinski
3. Analysis of Probabilistic Signals and Systems, Fall 2017, Professor James Matthews
4. Principles of Communication Systems, Fall 2017, Professor Alexander Wyglinski
5. Optimal Control, Spring 2017, Professor Raghvendra Cowlagi

Template: Courtesy of Jon Barron.