Meghanath Macha

I architect behavioral foundation models and the platforms that run them in production.

At ZeroToOne.AI, I lead AI and platform. We train a spatiotemporal model on where people go and what they do online to predict visits and intent. We score billions of IDs a day. I design the inference, identity, and data systems that serve those predictions and lead the teams integrating GroundTruth and WeatherBug.

Head of AI & Platform, ZeroToOne.AI

Meghanath Macha

Selected work

01

One model, many products

We replaced single-purpose audience classifiers with a transformer-based behavioral model trained on sequences of physical and digital events. It powers daily predictive audiences and provides a shared representation for warehouse enrichment and downstream models for churn, customer value, and next best action. A retailer can buy cheap clicks from people who never visit its stores. We keep bidding separate so campaign incentives do not change what a visit score predicts. In March 2026, Horizon Media reported 48% greater visitation efficiency versus baseline in one pilot combining our predictions with its own audience intelligence and campaign design.

We separate GPU embedding generation from CPU audience scoring. Once the embeddings are ready, daily scoring can run without waiting for GPU capacity. Automatic promotion requires each weekly retrain to clear ranking and lift thresholds across category levels and prediction horizons. Our ML platform tracks which features and datasets each model version depends on.

02

Getting AI to follow brand rules

A draft can follow the tone of a brand prompt and still miss a required trademark symbol. At Adobe GenStudio, I built the verification layer. We encoded brand guidelines as structured rules in a knowledge graph, including where each rule applied. After generation, GenStudio checked the draft against those rules and showed reviewers which guideline it failed and why. They could revise the asset and rerun the check before approval. The verification layer shipped in October 2024 with GenStudio for Performance Marketing.

03

Audiences you can describe and check

Custom audience requests used to sit with engineers for about two weeks. I built Audience Builder so a planner could describe who they wanted to reach and have the agent compile the brief into a typed filter specification over our model’s predictions. An answer can sound reasonable while dropping a constraint like a time window or geography. Planners can inspect and correct the filters before activation and reuse the definition. We validate the category codes and time windows in the compiled filters, then hash the definition to verify that the saved filters match the ones we sized. In GroundTruth Ads Manager, custom audiences are available within 24 hours.

04

Deciding what to buy and what to build

Ahead of ZeroToOne.AI’s acquisition of GroundTruth and WeatherBug, I led technical diligence on platform capabilities, the cost of rebuilding them, and overlap in data and identity systems. I now lead the integration across two live ad platforms. We moved legacy custom audiences to Audience Builder, consolidated GroundTruth’s direct audience delivery, and retired redundant compute pipelines while preserving audience IDs for live campaigns. We are also reconciling overlapping device identifiers across data sources. In GroundTruth Ads Manager, we expanded the predictive catalog from 40 audiences at launch to more than 1,500. Advertisers can compare those segments with standard visitation audiences in the same campaign workflow.

Notes

  • Where did the ML research go?

    Training loops are becoming routine engineering; the real ML research moved into ungameable verifiers and reward integrity.

  • Your synthetic customer should have a past

    Simulating customers from observed behavior makes predictions you can check against real outcomes and past experiments.

  • The harness is a mold, not a product

    Keep cognitive scaffolding replaceable, keep system boundaries hard, and use verified traces to train smaller models when volume justifies it.

More notes

About

I have taken models from research into production at the Carnegie Mellon spin-out that became ZeroToOne.AI, as a founding engineer on Adobe GenStudio, and again after returning to ZeroToOne.AI to build the Large Behavioral Model. I have a PhD in Information Systems from Carnegie Mellon, where I worked on privacy-preserving models of location data. My current research asks how much a model trained on sparse location histories can tell us about a future visit beyond what the most recent visit already tells us. I have published in ISR, KDD, DMKD, and Management Science.

I design systems so people can inspect what they are about to use. That means keeping audience filters readable, checking brand rules separately from generation, and keeping the meaning of an intent score consistent across products.

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

Contact

Architecture advisory. Technical diligence. Speaking.