Lovable is one of the fastest-growing AI companies, and millions of people use it to build software with natural language. Under the hood, autonomous agents generate code, execute commands, access tools, and take actions for users in real time.
“White Circle protects millions of AI interactions every day while helping us move faster.”
As a leading AI-first company, you run into a problem most companies haven't had to solve yet: controlling what an autonomous AI system is allowed to do in production, across millions of sessions, against constantly changing attack patterns.
Problem
Users attack the system. People try to generate phishing websites, explicit content, malware, gambling sites, and other types of inappropriate content. They try to run unsafe system commands, and attempt to manipulate the agent during execution - some attacks are simple, others are multi-step and designed to bypass safeguards.
Models themselves are also a source of risk. AI models have already pushed users toward suicide, sold products for $1, wiped email inboxes, erased entire hard drives, and gone rogue during training to mine crypto on GPUs - and in an agentic environment, this sounds scary. Lovable needs a system that adapts to both user behavior and model behavior continuously, without breaking the experience for normal users.
Integration
White Circle acts as a supervision layer in Lovable's pipeline. It's model-agnostic, and it operates independently from the underlying providers and models. Lovable sends every new event in a session to a single API endpoint, and White Circle evaluates and enforces policies in real-time.
White Circle covers:
- User supervision — catching attempts to override system instructions, manipulate agent behavior, generate harmful or illegal content
- Model supervision — monitoring what the model actually does
- Abuse and anomaly detection — identifying abuse patterns across sessions
- User risk scoring — building risk profiles at the user level so repeat offenders can be tracked across sessions
- Custom moderation policies — tuned for Lovable's product context, adjustable per feature and per use case
How this works
Most companies that try to build moderation in-house follow the same path: curate a dataset, train a single-purpose classifier, evaluate it, deploy it. For one policy, that process takes months. For hundreds of policies across different features and use cases, it simply doesn't scale, and it doesn't keep up with how fast models, attacks, and product requirements change.
White Circle uses its own proprietary self-adjusting classifiers. Lovable can define a new policy, test it against historical traffic using backtesting tools, and push it to production - without training dedicated models for every policy and waiting months for coverage. Policies can be scoped per feature, per use case, per type of event, or per user segment.
Lovable can switch providers or upgrade models without rebuilding its supervision layer because White Circle operates independently from the model stack.
Latency
Lovable's product is a conversational interface where people build software. Every millisecond of added latency hurts the experience, and a safety layer that makes the product feel slower eventually gets turned off. White Circle was designed to have very low latency to be used in realtime - in practice, the system is effectively invisible to normal users.
Tooling
White Circle gives Lovable operational visibility into model and user behavior across the platform:
- Dashboard with real-time metrics
- Alerts for anomalous patterns and emerging attack vectors
- Audit logs for compliance and incident review
- Tools for testing policy coverage before rolling out changes
- Human review queue for edge cases that need manual judgment
- User-level risk scoring to track persistent bad actors over time
- Analytics for tracking user sentiment, failure patterns, and common friction points