Sutanu Bhowmick

Agentic AI Architect · The Home Depot · PhD, Rice University · Washington, DC

I build AI systems at enterprise scale and own them end to end: architecture, deployment, and the monitoring that keeps them running afterwards. What I am measured on is the business result, not the model.

At The Home Depot I lead agentic AI initiatives. That means multi-agent pipelines built on frontier models, running multi-step workflows with little human intervention. My work also covers multimodal AI, computer vision, and MLOps.

Nine years in, the pattern I keep seeing is that the model is rarely what decides a project. Evaluation, cost, latency, and what happens when something fails decide it. That is where the time goes, and it is where the results come from: cost reduced by automating work that was being done by hand, multimodal systems that beat the baselines they were measured against, and pipelines that stay up under real load.

What I work on

Agentic AI Multi-agent pipelines on frontier models, handling multi-step workflows that used to need a person at every step.
Multimodal AI and computer vision Ten years of building systems that pull measurements and decisions out of images and video.
MLOps and scale Deployment, monitoring, cost, and reliability. The part that determines whether a system is still useful a year later.
Physics-informed machine learning Estimating the parameters of physical models from noisy sensor data. This is where SNAPE comes from.

SNAPE

SNAPE is a method I developed during my PhD for a problem that shows up everywhere in engineering and science. You have a physical model, you have noisy measurements, and you need the model's parameters.

The standard approach differentiates the measurements numerically. Noise makes that unreliable, and the higher the derivative, the worse it gets, so published methods start failing at a few percent of noise. SNAPE never differentiates the data. It fits smooth basis functions to the measurements and solves for the physical parameters at the same time, so every derivative comes from the fitted function instead of the raw signal. It also does not need the initial or boundary conditions, and it reports uncertainty bounds alongside the parameters.

It recovers the parameters of Burgers' equation from data carrying 100% Gaussian noise, and identifies a fifth-order Euler–Bernoulli beam model from ordinary video of a vibrating beam. It applies anywhere a known physical model has to be calibrated from noisy sensor data: structural and machine digital twins, medical imaging, power grid dynamics, thermal testing.

The connection to the rest of my work is not a coincidence. SNAPE exists because deep learning was not repeatable enough for this problem. Knowing when a model is wrong, and how much to trust it when it is right, is the same discipline that decides whether an AI system is safe to run at scale.

An open-source implementation, pysnape, is available at sutanubhowmick.github.io/pysnape.

Selected research

Over 1,500 citations, mostly in structural health monitoring, computer vision for civil infrastructure, and parameter estimation.

Data- and theory-guided learning of partial differential equations using SimultaNeous basis function Approximation and Parameter Estimation (SNAPE), Mechanical Systems and Signal Processing 189 (2023) 110059. 10.1016/j.ymssp.2022.110059
Physics-guided identification of Euler–Bernoulli beam PDE model from full-field displacement response with SNAPE, Engineering Structures 289 (2023) 116231. 10.1016/j.engstruct.2023.116231
Review of bridge structural health monitoring aided by big data and artificial intelligence: from condition assessment to damage detection, Journal of Structural Engineering 146 (2020). 10.1061/(ASCE)ST.1943-541X.0002535
Measurement of full-field displacement time history of a vibrating continuous edge from video, Mechanical Systems and Signal Processing 144 (2020) 106847. 10.1016/j.ymssp.2020.106847

Full publication list on Google Scholar