Predicting Churn and Catching Bad Data
EdTech / Data OperationsPredicting which users are about to churn and catching invalid events before they reach downstream analytics, for an EdTech platform and a high-frequency data operation.
Overview
A consumer EdTech platform needed to know which users were about to churn before they did, and a separate high-frequency data operation needed to catch invalid or fraudulent events before they polluted downstream analytics. Canvas Labs Partners built a churn-prediction model for the first client and an ML-based anomaly-detection layer for the second, each tuned to the data it runs on.
2.5x user retention
89% recall, 75% precision from the churn-prediction model
92% recall
catching invalid and anomalous events at ingestion
Challenge
One client needed to know which users were about to churn before they did, so retention strategy could act on a prediction rather than a departure. Another needed to catch invalid or fraudulent events inside a high-frequency data pipeline before they polluted downstream analytics, which meant validating events at the moment they arrived rather than after the fact.
Solution
For the consumer EdTech platform, Canvas Labs Partners built a cohort and churn-prediction model that identifies power users and at-risk segments directly from behavioral data. The model feeds straight into retention strategy, so the team sees which users are worth intervening on before a prediction becomes a cancellation.
For the high-frequency data operation, Canvas Labs Partners designed an ML-based anomaly-detection layer running on GCP BigQuery, validating event data the moment it is ingested rather than after it has already reached downstream analytics. Built in Python, the layer flags invalid or fraudulent events at the point of entry, before they reach a single report.
Both systems share the same structure: a model trained to flag risk before it becomes a problem, rather than a report that explains it afterward. The churn model identifies a departure before it happens; the anomaly-detection layer catches a bad event before it ever reaches a dashboard. Predictive modeling sits underneath both, tuned to the signal each client needed to act on early.
Results & benefits
The churn-prediction model drove a 2.5x spike in user retention, running at 89 percent recall and 75 percent precision, so the EdTech platform's retention strategy now acts on users genuinely at risk. The anomaly-detection layer on the high-frequency data operation's pipeline catches invalid and anomalous events at 92 percent recall, before they reach downstream analytics. Both systems turn a reactive problem into a predictive one: a departing user or a bad event is flagged ahead of time, by a model tuned to that specific client's data, rather than found after the fact.
Customer
EdTech platform & high-frequency data operation
Services
- Agentic AI consulting & development
- Predictive modeling & ML engineering
- Real-time data validation
Tech stack
About us
Canvas Labs Partners is a software consulting team that designs and ships production agentic AI systems: multi-agent pipelines, tool-calling architectures, and human-governed automation, together with the trading infrastructure, blockchain, mobile, and product engineering those agents plug into. We publish our work as a case ledger rather than a portfolio, so every claim on this site links to the engagement behind it. Engagements follow the same four-step discipline, Map, Bound, Instrument and Govern, whatever the stack, and every entry in the ledger carries a result we are willing to be measured against.
Delivered individually, in a prior role, by one of Canvas Labs Partners’ founding team members. Not by the firm itself.
Have a system like this one?
We built this one end to end. If you are facing something similar in edtech, book a free discovery call and we will tell you what it would take.