The developer of an AI-based Telegram news aggregator faced technical constraints due to third-party service limits. A free RSS conversion service restricts requests to 13-15 per few minutes, forcing a 75-second pause between topics and a 40-45 minute cycle for 12 topics.
To manage automation platform costs, the developer replaced individual schedules for 12 topics with a single 'Orchestrator' entry point. This prevents exceeding the monthly limit of 2,500 runs and the limit of 5 concurrent executions, as internal calls are not counted.
Initial deduplication failed because cluster_id, based on the first 200 characters of text, could not identify identical events described differently. A new experimental mechanism using headline comparison achieved 89% accuracy, though it is not yet integrated into the main feed.
Redundant speech synthesis accounted for 56% of costs, with 2,275 calls for 1,003 unique items. By switching to sequential processing via the Orchestrator, the developer enabled a cache that prevents re-processing items already synthesized within 48 hours.
Increasing the collection frequency to once per hour exhausted the daily 500-request quota for the AI classifier within the first hour. The system failed to alert the developer because the email notification system was disabled, leaving the failure unnoticed without manual checks.
The project is in early beta with 1,037 listens recorded in August. While external user growth is minimal, retention remains between 60% and 80%. The developer plans to implement batch processing for AI calls to optimize costs and reduce redundant model requests.
