SceneIt explores a practical architecture question: how should a product combine public catalog data, authenticated preferences, natural-language discovery, and generative explanations without coupling the UI directly to every provider?
The browser talks to a stable application API rather than TMDB or Gemini directly. Next.js route handlers keep credentials server-side, allow-list upstream query parameters, normalize errors, and apply a 30-minute revalidation policy. Supabase is isolated behind SSR-aware clients and session middleware.
Decision 01 / Integration
All catalog calls pass through typed server utilities and application-owned routes. Discover requests forward only known filters, and upstream failures become consistent API responses.
Why it matters: credentials stay private, provider details do not leak across UI components, and the integration can evolve in one place.
Decision 02 / Recommendation quality
Natural language is parsed into explicit genres, years, rating, runtime, provider, region, and mood constraints. Results are fetched, enriched, filtered again, and returned with the applied plan visible to the user.
Tradeoff: narrower language coverage than an unconstrained agent, but predictable recommendations that can be debugged and tested.
Decision 03 / AI safety
Recap requests first resolve the title and build context from TMDB. Gemini receives that source text with an instruction not to invent missing detail. If the model or key is unavailable, the route returns the source summary.
Why it matters: the useful path survives a model outage, and generated answers have an explicit evidence boundary.
Decision 04 / Product access
Core browsing and detail routes work without an account. Middleware refreshes sessions and protects profile, favorites, and onboarding surfaces, preserving a low-friction evaluation path.
Why it matters: authentication supports the product instead of becoming a gate in front of its primary value.
01
Next.js revalidates TMDB responses every 30 minutes, while TanStack Query deduplicates client requests and manages loading, error, and stale states.
02
Broad discovery reads one result page. Narrow prompts search deeper, then enforce constraints locally. The extra calls are spent only where precision needs them.
03
Runtime and provider lookups run concurrently where possible, and optional enrichment failures degrade individual results rather than failing the complete response.
I intentionally present SceneIt as a product prototype, not a production system. The senior-level work is visible in the boundaries already chosen and in identifying the controls required before serving real users at scale.
