What it is
An engine that runs an affiliate content business end to end: it discovers products worth recommending, writes content shaped for each platform, renders images and short videos, publishes them, and measures what converts. It is operated from a Telegram bot and drives traffic to a public niche website (linked above).
How it's built
- Six-stage worker pipeline — hunt → generate → render → publish → measure → reap. Workers never call each other; they coordinate through MongoDB state machines with atomic claiming, so every stage is resumable and idempotent.
- AI only where judgment is needed — agents score products and write scripts, with Zod-validated structured output; platform-specialist agents (Pinterest, Instagram) and an evaluator/feedback loop improve results over time.
- Compliance is code, never an LLM — a deterministic gate enforces affiliate disclosure and blocks fake urgency or invented claims.
- Safety governor — token buckets, posting windows, jitter, gradual ramp-up, similarity checks and shadow-ban detection keep accounts well below platform limits.
- Media pipeline — images with sharp, videos with FFmpeg, Edge-TTS voiceovers, captions and stock B-roll.
- Operations — encrypted platform tokens (AES-256-GCM), Docker and PM2, GitHub Actions CI/CD with health checks, a 34-file Vitest suite, and LLM calls through a self-hosted FreeLLMAPI gateway.
Stack
TypeScript, Node.js, Express, MongoDB, grammY (Telegram), Zod, sharp, FFmpeg, Edge-TTS, Next.js, Docker, PM2, GitHub Actions.
Source code is private.