01
Prediction and decisioning
Build calibrated prediction across scenarios and production decisioning with simulation and optimization under uncertainty.
Staff Applied Scientist & Tech Lead, Applied AI
Tao Cao
Applied AI tech lead building production prediction, decisioning, and agentic systems, with a focus on simulation, evaluation, and physical AI.
Systems
Prediction and decisioning from 0-to-1 to enterprise scale
Science
Deep learning, simulation, optimization
Agents
Planning, tool use, RCA, evaluation
Lead
Architecture, roadmaps, teams, stakeholders
Professional focus
Production AI systems and technical leadership
At Walmart, I lead two production systems: the Delivery ETA Engine and Driver Dispatch. The work spans unified prediction, simulation-backed decisioning optimization, agentic model RCA, evaluation, training, and serving.
I own architecture and technical roadmaps while leading 5-10 DS, DE, and MLE contributors. These systems have delivered significant business impact through growth and cost savings, with recognition through KDD 2024, a patent filing, and two Walmart AI Summit finalist selections.
02
Agents and evaluation
Develop multi-agent planning, tool execution, model RCA, and evaluation workflows with traceable state and outcomes.
03
Production and leadership
Scale distributed training, PySpark pipelines, and online and asynchronous/batch inference while leading architecture, roadmaps, and teams.
Independent work
Selected prototypes
Practical tools for decisions that are hard to structure, compare, or explain.
AI Travel Planning Assistant
A representative project from my broader agentic portfolio. It plans the messy middle of travel—what to do and where to eat—using external search, ranking, joint schedule optimization, deterministic evaluation, and traceable execution.
Three bounded agents divide interaction, structured trip intake, and reasoning/tool execution. Deterministic services own versioned state, validation, approval, and run status so voice and visual review remain aligned.
Multi-agent planning / Tool execution / Search / Optimization / Evals
01
Voice agent
Natural interaction and turn-taking
02
Profile agent
Schema-grounded trip intake
03
Planning agent
Search, optimize, evaluate, execute
Notes from building
Writing and talks
-
Build note
Adding a Reliable Voice Layer to the AI Travel Planner
Why three bounded agents, schema contracts, and deterministic state ownership made voice intake more reliable.
-
Build note
I Built an AI Travel Planner for the Messy Middle of Trip Planning
Why bounded workflows, grounding, optimization, evaluation, and traces matter in an agentic product.
-
Build note
Two Mortgage Calculators I Built to Make Prepayment Decisions Clearer
Making loan strategy tradeoffs visible through cash-flow timing, NPV, and actual payment history.
-
Build note
Turning Dull Manuals into Actionable Maintenance
A grounded multi-agent workflow for extracting, refining, and prioritizing appliance maintenance tasks.
-
Analysis
Your Heat Pump Might Be Wasting Energy Because of Your Thermostat
Quantifying a compatibility gap between variable-speed equipment and common smart thermostats.
-
Talk
Driver Search Decision-Making at KDD 2024
A TSMO Workshop talk on building large-scale AI decision systems for crowdsourced delivery.
Background
Applied AI for software and physical systems
My foundation in modeling and optimization shapes how I build production AI across prediction, decisioning, simulation, agents, and evaluation. Earlier work in energy, digital twins, and engineering software grounds my growing focus on physical AI.
I stay close to the technical work: architecture, model design, training, serving, evaluation, team direction, and stakeholder problem formulation.
- Ph.D.
- Mechanical Engineering, Modeling & Optimization
University of Maryland - M.S.
- Computer Science, Machine Learning
Georgia Institute of Technology - Domains
- Prediction, decisioning, agentic systems, and physical AI