Reasoning over data
Teaching agents to decompose ambiguous analytical questions, form hypotheses, sequence operations, and revise a plan as new evidence appears.
Applied AI · Agents · Data science
Building practical AI for complex work.
I’m Subhankar, a Senior Research Scientist at NVIDIA. I build AI agents that help people navigate data science tasks and turn complex analytical workflows into more intuitive experiences.
NVIDIAGoogleMicrosoftUIUCNIT Rourkela
Current focus
Making complex data work more intuitive.
My current work focuses on reasoning: helping models break open-ended data science questions into defensible steps, choose and use the right tools, inspect intermediate results, and adapt when an approach does not work.
Tool-using agents
Building models that can select analytical tools, work across multi-step data science workflows, and connect each action to the task’s underlying intent.
Grounded assistance
Making agent behavior easier to trust through visible assumptions, intermediate checks, recoverable errors, and results grounded in the available data.
Previous research
Selected work in speech AI.
MagpieTTS-LF
Generating coherent long-form speech at inference time—without retraining on long-form data.
View on ScholarLearning monotonic alignment
Improving the reliability of LLM-based speech synthesis by reducing repeated, missing, and misaligned speech.
View on ScholarKoel-TTS
Aligning speech generation with intelligibility, speaker identity, and transcript adherence.
View on ScholarA statistical foundation
From computer science and statistics to applied AI.
My work is shaped by a B.Tech. in Computer Science & Engineering from NIT Rourkela and an M.S. in Statistics from the University of Illinois Urbana-Champaign.
More about my pathLet’s connect