How Prescient Uses Machine Learning and Modelbit to Land Billion-Dollar Customers
How Prescient Uses Machine Learning and Modelbit to Land Billion-Dollar Customers
by
Cody Creco
,
Founder and CTO
A cutting-edge approach to machine-led measurement
Prescient is the first machine learning-powered marketing attribution platform. Our founders used Mixed Media Modeling and cutting-edge research in causal modeling to build the first successful marketing attribution platform for streaming music. When Apple upended the industry by blocking streaming pixels on iPhones, we were well-positioned to bring our pixel-free, AI-based platform to the eCommerce market.
To build a full attribution model for each customer, we ETL customers’ spend data and session-level data into per-customer datamarts in Snowflake. From there, we can build a full causal model of the customer’s marketing spend, halo effects, and revenue results.
Forecasting based on proposed changes
Our top request from customers and prospects is to forecast what would happen if they made changes to their ad spend. Before making big changes, our customers want confidence in understanding what the results will be ahead of time.
To build these forecast models, we train one model for each marketing campaign for each customer. We use Tensorflow to train neural nets that can accurately represent the patterns in each customer’s unique dataset, no matter how complex those patterns might be. These neural nets can accurately represent the local maximum you might get from certain ad networks when you hit a certain ad spend threshold; and the synergistic benefits customers get from using certain combinations of ad networks.
Continuous deployment using Modelbit
We retrain and redeploy thousands of models daily, so choosing the right partner for model deployment is critical. We initially investigated Seldon and SageMaker, but found them too complex and too rigid. Modelbit’s performance paired with its ease of use was a game-changer for us.
Each day, a job kicks off for each marketing campaign at each customer, retraining their models on the latest data. When the training completes, we use the Modelbit API to add the latest models to our Modelbit deployment.
In our product, customers use an easy web interface to simulate changes to their spend and see live predictions of how their revenue would be impacted. This simulator makes live calls to Modelbit’s REST API to call the most recently-deployed version of that customers models.
Modelbit was incredibly fast to prototype with during the prototyping phase, and has proved stable and reliable now that we are running thousands of models in production. Modelbit handles all the DevOps, security and stability-related work so that we can safely deploy new models to production without requiring the involvement of the software engineering team.
Billion-dollar impact
By using Modelbit to deploy the models that solve our top customer request, we’ve built our strongest differentiator. The forecasting model has become our silver bullet. None of our competitors have anything close to it. And by adding it to our product, we were able to onboard two $1B companies as customers – our biggest customers to date.
A cutting-edge approach to machine-led measurement
Prescient is the first machine learning-powered marketing attribution platform. Our founders used Mixed Media Modeling and cutting-edge research in causal modeling to build the first successful marketing attribution platform for streaming music. When Apple upended the industry by blocking streaming pixels on iPhones, we were well-positioned to bring our pixel-free, AI-based platform to the eCommerce market.
To build a full attribution model for each customer, we ETL customers’ spend data and session-level data into per-customer datamarts in Snowflake. From there, we can build a full causal model of the customer’s marketing spend, halo effects, and revenue results.
Forecasting based on proposed changes
Our top request from customers and prospects is to forecast what would happen if they made changes to their ad spend. Before making big changes, our customers want confidence in understanding what the results will be ahead of time.
To build these forecast models, we train one model for each marketing campaign for each customer. We use Tensorflow to train neural nets that can accurately represent the patterns in each customer’s unique dataset, no matter how complex those patterns might be. These neural nets can accurately represent the local maximum you might get from certain ad networks when you hit a certain ad spend threshold; and the synergistic benefits customers get from using certain combinations of ad networks.
Continuous deployment using Modelbit
We retrain and redeploy thousands of models daily, so choosing the right partner for model deployment is critical. We initially investigated Seldon and SageMaker, but found them too complex and too rigid. Modelbit’s performance paired with its ease of use was a game-changer for us.
Each day, a job kicks off for each marketing campaign at each customer, retraining their models on the latest data. When the training completes, we use the Modelbit API to add the latest models to our Modelbit deployment.
In our product, customers use an easy web interface to simulate changes to their spend and see live predictions of how their revenue would be impacted. This simulator makes live calls to Modelbit’s REST API to call the most recently-deployed version of that customers models.
Modelbit was incredibly fast to prototype with during the prototyping phase, and has proved stable and reliable now that we are running thousands of models in production. Modelbit handles all the DevOps, security and stability-related work so that we can safely deploy new models to production without requiring the involvement of the software engineering team.
Billion-dollar impact
By using Modelbit to deploy the models that solve our top customer request, we’ve built our strongest differentiator. The forecasting model has become our silver bullet. None of our competitors have anything close to it. And by adding it to our product, we were able to onboard two $1B companies as customers – our biggest customers to date.