Setup & env vars

Run the dev server

cd /Users/intern-vidhan/Documents/motion-saas-demo
npm install
npm run dev

App opens at http://localhost:3000. Hot reload via Turbopack.

Environment variables

All env reads happen server-side in API routes and lib/streamGenerate.ts. Put them in .env.local (gitignored).

Required for non-mock generation

# Either of these gets you off the mock generator:
GEMINI_API_KEY=AIza...            # https://aistudio.google.com/apikey — free tier
NVIDIA_NIM_API_KEY=nvapi-...      # https://build.nvidia.com — 1000 free credits

If neither is set, the mock generator runs. The app still works, scenes still render, but the copy is templated rather than AI-written.

Optional model overrides

GEMINI_MODEL=gemini-2.5-flash-lite      # default — fastest, ~2.3s/call
# GEMINI_MODEL=gemini-2.5-flash         # better copy, ~6-8s/call
# GEMINI_MODEL=gemini-3-flash-preview   # newest but rate-limited on free tier

NVIDIA_NIM_MODEL=google/gemma-4-31b-it  # default fallback
# Some other useful NIM models:
# meta/llama-3.3-70b-instruct           # fast (~4s), good quality
# nvidia/llama-3.3-nemotron-super-49b-v1.5  # slow but reasoning-capable

Timeouts (rarely needed)

NVIDIA_NIM_TIMEOUT_MS=15000   # cold-start tolerance on NIM, default 15s
GEMINI_TIMEOUT_MS=20000       # Gemini per-call cap, default 20s

Free-tier reality check

Gemini's free tier on gemini-2.5-flash-lite allows roughly:

  • 1,500 RPD (requests per day)
  • 15 RPM (requests per minute)
  • 1M TPM (tokens per minute)

One ad generation = 1 Director call + N Specialist calls (5–6 per ad) = ~7 requests per ad. That's ~200 ads/day before exhausting RPD. Plenty for prototyping.

When you exceed the free tier:

  • Paid Gemini 2.5 Flash Lite: $0.10 / 1M input tokens, $0.40 / 1M output. ~$0.001 per ad. Effectively free.
  • Or fall through to NIM (separate quota, also free up to 1k credits).

Reset state if quota is exhausted

If /api/generate returns 429 (rate limit) repeatedly:

  1. Switch model: GEMINI_MODEL=gemini-2.5-flash-lite → gemini-2.5-flash (separate per-model bucket)
  2. Or switch provider: comment out GEMINI_API_KEY so it falls through to NIM
  3. Or wait — RPD resets at 00:00 PT each day (12:30 PM IST)

Production deployment

Not done yet. The app works on Vercel out of the box (it's just Next.js), but you'll need to:

  • Move env vars to Vercel's project settings
  • Decide on render strategy (Remotion <Player> in browser is fine for preview; MP4 export requires Remotion Lambda)
  • Likely add billing/auth — currently none