The 1trAIner service employs a cascading planning architecture to ensure that AI-generated training plans remain consistent with long-term goals. Instead of generating daily tasks independently, the system builds them upon weekly, monthly, and seasonal contexts. If a required upper-level plan is missing, the service generates it before creating the daily schedule.
To ensure quality, the system enforces a minimum length of 300 characters for any generated plan. If a response is shorter, it is treated as a failure. Additionally, the service integrates external data, including Garmin activity metrics, recovery indicators like heart rate variability, and user-defined constraints such as blocked dates, to adjust training intensity.
The platform optimizes costs by scheduling background tasks during off-peak hours. As of September 10, 2026, standard input costs $0.15 per million tokens, while cached input is $0.003. During peak hours (01:00–04:00 and 06:00–10:00 UTC on weekdays), these rates double. Consequently, daily plans are generated between 03:15 and 03:59 Moscow time to leverage lower pricing.
The development team implemented several safeguards to prevent common AI errors. These include validating the existence of a parent plan, filtering out placeholder responses, and using domain-specific prompts for voice message transcription via faster-whisper. Future efforts will focus on expanding automated testing for plan structure and adherence to user constraints.
