Meghanath Macha.

Meghanath Macha

I build models of consumer behavior, and the systems and teams that turn them into products.

I started out training the models and still design the systems. At ZeroToOne.AI, I lead AI and platform: we train one shared model on physical and digital behavior, then use it to predict visits, intent and other outcomes. Much of my work now is the data and infrastructure behind it.

Head of AI & Platform, ZeroToOne.AI

Meghanath Macha

About

Three times I have taken a model from research into production: at the Carnegie Mellon spin-out that became ZeroToOne.AI, at Adobe GenStudio, and back at ZeroToOne.AI with the Large Behavioral Model.

I led the technical diligence for GroundTruth and WeatherBug and now lead the integration of the platforms and teams. Before that, I was a founding engineer on Adobe GenStudio, where I built the layer that keeps GenAI content on-brand, grounding generation in a knowledge graph of the brand's guidelines rather than in a prompt.

I have a PhD from Carnegie Mellon, where I worked on privacy-preserving models of location data. Papers in ISR, KDD, DMKD and Management Science.

Selected work

01

One model, many products

We used to train a separate classifier for each audience. Now one vertical foundation model, trained on sequences of physical and digital behavior and scoring 1.7B+ IDs a day, is the base layer: predictive audiences refreshed daily and delivered inside the ad platform, enrichment feeds for a client's own data warehouse, and managed models for churn, customer value and next best action. Horizon Media, which runs the model inside HorizonOS, reported 48% better visitation efficiency in a recent pilot, and campaign outcomes flow back into training.

02

Getting AI to follow brand rules

At Adobe GenStudio, each brand rule was an object in a knowledge graph that generated content could be tested against. Verification ran as a separate step from generation, so the rules did not depend on which model wrote the content, and a reviewer could see which rule an asset failed. It was the basis for a fully AI-generated marketing campaign and cut content production cycle time by 60%.

03

Audiences you can describe and check

Audience requests used to go through engineers and took about two weeks. Now a planner describes who they want to reach, the agent turns that into an audience definition over the model's predictions, and the planner reviews what the system understood before using it. Time to a new audience is 24 to 48 hours, set by the daily scoring refresh rather than by the agent. The definition is a readable object rather than a trained artifact, so it can be checked, compared and reused across products.

04

Deciding what to buy and what to build

Before ZeroToOne.AI acquired GroundTruth and WeatherBug, I led the technical diligence: what each platform could do, what would have to be rebuilt, and where their data and identity systems overlapped with ours. The integration plan that followed had to keep two live ad platforms running while the models, data layer and teams were brought together. That is the work I lead now.

Background

ZeroToOne.AI · Head of AI & Platform
Adobe · GenStudio / Machine Learning
ZeroToOne.AI, then a Carnegie Mellon spin-out · Founding ML Engineer
Adobe Research · Research Scientist
Carnegie Mellon University

PhD in Information Systems

IIT Kharagpur

BSc + MSc in Mathematics & Computing

Research

10 granted patents, 15+ filedInformation Systems Research · KDD · DMKD · Management Science
Explaining anomalies in groups with characterizing subspace rulesDMKD · 2018
Personalized privacy preservation in consumer mobile trajectoriesISR · 2023
Privacy choice during crisisManagement Science · 2026
Revisiting how to focus: Triplet attention for joint entity and relation extractionDI @ KDD · 2022
Modeling a user predisposition based on location dataPatent · ZeroToOne.AI · 2025
All publications on Google Scholar

Notes

More notes

Contact

If something here interests you, write to me.

Advisory. Speaking.

Email LinkedIn Google Scholar ZTO engineering blog