Building a Practical AI Radar — notes from the state-management trenches

Short summary

When building Radarix.ai, state management through SQLite became the critical bottleneck, not browser automation. The architecture separates a control plane (holding DB and cron) from two worker VPSes running Playwright in Docker, allowing fault isolation and concurrent work. The key insight: use models as evidence reviewers to validate browser actions rather than as drivers, avoiding hallucinations and maintaining registry integrity.

• •SQLite as the single source of truth for submission state prevents rebuild disasters and data loss
• •Separate control plane from worker VPSes for fault isolation, IP diversification, and concurrent execution
• •Models as evidence reviewers (not action drivers) eliminates hallucinations; deterministic tools collect, models validate

Generated with AI, which can make mistakes.