Moloco’s CARA
Compound Ad Recommendation Architecture for Moloco Ads
Overview
CARA (Compound Ad Recommendation Architecture) is Moloco’s AI system that orchestrates more than 30 models, agents, and specialized software to deliver performance outcomes on the open Internet.
Advertising on the open Internet is complex. The ecosystem spans billions of users, over a trillion daily ad opportunities, thousands of advertisers, and millions of apps. Every decision needs to be at the ad impression level and needs to be made within tens of milliseconds. Most advertiser-user-context combinations have never appeared before. And the signals AI draws from arrive in different formats, at different speeds, and with different levels of reliability.
Moloco has been building for these unique complexities since 2013, long before the current wave of AI excitement. CARA uses deep learning architectures to ingest signals from many sources at once and form critical insights on how users, apps, creatives, and conversions relate. Across more than one trillion ad opportunities per day on average, CARA estimates user intent, conversion probability, creative fit, and bid economics – uniquely for each advertiser, in real-time. ⓘ
Architecture
For Moloco Ads, the system is organized into six technical domains, composed of multiple models, agents, and specialized software: Campaign Automation, Supply, Ad Recommendations, Bidding, Creative, and Signals.
Domain details
Each domain solves a distinct challenge, and works in concert to achieve profitable performance outcomes.
Advertisers input what success looks like for their business: maximizing return on ad spend or driving in-app events such as registrations, deposits, or purchases. Additionally, CARA agents evaluate live campaigns, surface opportunities to improve performance, and the resulting insights to advertisers.
For every impression, CARA estimates the value of the opportunity based on how likely a user is to take a particular action, measured against the value of that action to that advertiser. To be able to do this, our models use deep learning to ingest signals from many sources simultaneously, forming critical insights of user behaviors, publisher apps, products, creative formats, content, and the interaction among them.
Every user sees a creative chosen specifically for them: Creative is a per-impression decision, not a campaign-level assessment. For the Moloco SDK, CARA models tune individual ad components that the SDK assembles into full ad units at impression time. CARA's selection models then identify the combination most likely to perform for that user, in that context, from the vast candidate space.
Signals is the foundation for the other five domains. CARA ingests signals from a wide variety of advertiser conversion events, publisher interactions, bid requests, mobile measurement partners, and SDK activity. These signals arrive in different formats, at different speeds, and with different levels of reliability. Our system is built to integrate all of these signals and find the patterns that matter. Our models learn relationships across users, products, creatives, publisher contexts, and conversion events, forming a unified layer that makes every other domain sharper.
How the system compounds
Improvements in each domain strengthen the overall performance of the system and compound on one another. All domains feed into and read from a unified Signals layer, so as that domain becomes richer, every other domain benefits, even without a domain-specific change. The gains carry from one domain to the next: better Creative selection improves the training signal for Ad Recommendations, richer Supply signals improve Bidding accuracy, and more precise Bidding generates more outcome data that sharpens predictions downstream.
Our teams iterate in parallel to deliver performance improvements, running hundreds of concurrent experiments every day to test new approaches against live traffic. Those experiments contribute to model improvements, amounting to 65 model updates in 2025.
FAQ
A compound AI system coordinates multiple models, agents, and specialized software rather than relying on a single model. Each component in CARA is built and optimized for a specific domain: Bidding, Creative, Supply, and so on. The compound architecture means improvements in any domain propagate to others: a better Signals layer sharpens Bidding accuracy; better Creative selection produces richer training data for Ad Recommendations. These improvements multiply through the system rather than summing independently.
Updated model versions are deployed to production as often as every two hours. Real-world outcomes – such as conversions, clicks, and engagement signals – flow back continuously as training data. This continuous feedback loop means the system is always learning from the most recent data about what works for each advertiser, user segment, and supply context.
Sparse signal is one of the core challenges CARA was designed to address. On the open Internet, a meaningful share of inventory carries limited user signals, new users, low-activity app contexts, or environments where behavioral history is thin. CARA's sequence-modeling architecture (trained on user event histories across apps) in the Ad Recommendations domain is purpose-built to predict outcomes in these conditions, drawing on representations learned across the full breadth of Moloco's supply and advertiser relationships.
CARA represents Moloco's compound AI architecture, and the approach applies across Moloco's product suite. The specific technical domains and their configuration vary by product context. Information about how the compound AI architecture applies to Moloco Commerce Media can be found here.
The Moloco SDK serves two functions simultaneously. First, it integrates directly with publishers, so CARA accesses their ad inventory directly instead of routing through exchange partners. Second, it gives CARA direct control over creative rendering on devices, unlocking richer ad formats and behavioral signals back to the models. Each new SDK integration strengthens the entire compound system because the signal it generates improves training data quality across all six technical domains.
