Aura — event-driven media-analysis backend

An async, event-driven backend that ingests a social-media post URL, fetches its images, runs them through a multimodal LLM, and streams a short styled text with a structured tone breakdown back to the client in real time.

Built solo, by hand (summer 2025). Deployed live on AWS, later shelved when the upstream platform blocked datacenter-IP scraping. Backend only — the claimable substance is the architecture, not a product.

Architecture

• FastAPI producer — creates a Task, publishes to a Redis Stream, returns immediately
• Two dedicated workers (download + LLM) consuming via consumer groups (XGROUP/XREADGROUP/XACK), concurrency capped with asyncio.Semaphore
• Idempotency — pre-work DB check emits a skipped status when the image already exists
• Live status streaming — workers publish progress to a per-task stream (task:{id}:updates); an authenticated, ownership-checked WebSocket tails it with XREAD, terminating on done/failed/skipped
• Data model — eight SQLModel tables over PostgreSQL 15 with Alembic migrations; per-invocation LLM telemetry (model, token counts, latency) in its own table; tone breakdown as JSONB
• LLM integration — Gemini multimodal with response_mime_type="application/json", parsed into a Pydantic schema, tenacity exponential-backoff retries
• Ops — Docker Compose (postgres, redis, api, two workers, SMTP sink); GitHub Actions SSH-deploy to EC2 with a post-deploy health check