Can I vibecode Nsketch AI?

Build me a private AI media workbench to replace Nsketch AI. Requirements:

- A local web app on port 4317: Node.js 22, TypeScript, Fastify, React, and Vite.
  The server owns @fal-ai/client; the browser never receives the provider key.
- Implement exactly two workflows: text-to-image with fal-ai/nano-banana-2 and
  text-to-video with fal-ai/ltx-2.3/text-to-video/fast. Accept prompt and aspect ratio.
- Submit queued jobs server-side and save the provider request ID before polling.
  Resume unfinished jobs after restart; show queued, running, success, failed, and canceled.
- Store prompts, endpoints, parameters, state changes, cost estimates, errors, and local
  output paths in SQLite via better-sqlite3. Download files to media/YYYY-MM-DD/.
- The local page has a submission form and searchable gallery with compare, favorite,
  rerun, download, delete, and JSON metadata export actions.
- Read FAL_KEY from .env and ship .env.example. Validate inputs, use safe filenames,
  and show estimated per-job plus month-to-date API spend before a submission.
- Bind to localhost only. No accounts, billing, telemetry, or analytics; everything stays
  on my machine except prompts sent to fal.ai. Tests use a fake provider and spend nothing.
- Out of scope: the full model catalog, voice cloning, lip sync, motion transfer, social
  templates, hosted storage, public galleries, provider failover, and mobile apps.
- Include a README with setup, current endpoint costs, data and backup paths, provider
  safety policies, and a plain warning that every real generation spends money.