gora.
AI / generative media

GoraGen — AI video generation platform

Self-serve LoRA training, ComfyUI orchestration on rented GPUs, Gemini quality evaluation, multi-shot storyboards in one multi-tenant platform.

Region
goragen.com (global)
Role
Tech lead, architect, full-stack
Period
2025–2026
Industry
AI / generative media

Internal metrics not published. All details extracted from public README and repo.

Summary. Below — three parallel sections, each self-contained.

What's inside

  • Four apps behind one Caddy: landing (Astro), dash (React 19), admin (React 19), API (FastAPI).
  • Plan-100 Control Plane: coordinator + runner + reconciler with leader leases, six cutover stages, post-crash recovery.
  • 61 SQL migrations, Postgres-native; pg_advisory_lock(42) for multi-replica safety.
  • Isolated workers container (singleton, replicas:1, stop_grace 120s) with eight loops.
  • Iterative quality pipeline: RU→EN translation → Claude picks model → Vast.ai instance (40-80GB VRAM) → ComfyUI → Gemini scores on 4 criteria → decision (regenerate / acceptable / good).
  • Self-serve LoRA training: kohya_ss + BLIP-2 captioning + validation grid + automatic persona insert.
  • ChromaDB for semantic search of recipe prompts.
  • @idempotent middleware on 5 mutating endpoints (sha256 of body, 24h TTL, replay on retry).
  • Cookie JWT + Bearer gk_live_* API keys (Argon2 hashed).

Stack

  • FastAPI (Python)
  • psycopg v3
  • PostgreSQL 16
  • ChromaDB (vectors)
  • Cloudflare R2 (S3-compat)
  • Caddy
  • Docker compose
  • Vast.ai (GPU rental)
  • Kling Direct API
  • Anthropic Claude
  • Google Gemini 2.5 Pro
  • Sentry
  • ComfyUI
  • kohya_ss
  • BLIP-2
  • React 19
  • TanStack Router
  • Astro
  • Vite 6
  • Argon2