Data Flywheel: Ingest, Train, Serve on SQLite

A single-process service that accepts labeled data, persists it in SQLite, triggers model retraining after every N records using honker's durable job queue, and serves predictions from the latest model. The entire system — data, jobs, events, model metadata — lives in one .db file.

When to use this skill

• You need a self-improving service where incoming labeled data feeds back into model training
• You want async training that doesn't block the ingest or inference paths
• You want transactional guarantees — data persistence and training triggers commit or rollback atomically
• Your write throughput fits in a single process (thousands of ingests/sec is fine)
• You want to avoid adding Redis, RabbitMQ, or Celery for background jobs

When NOT to use this skill