Problem
Finding more titles is easy; choosing one that fits the moment, time and available services is harder.
Bilingual AI decision product · Functional local prototype
PlayNext
A movie and series decision assistant that turns mood, time and streaming constraints into a small, explainable shortlist grounded in real catalogue data.
Local prototype · Production deployment is not published yet
Target users
Viewers who want a small, trustworthy shortlist instead of another endless catalogue.
My contribution
Product strategy, UX, React frontend, API architecture, recommendation logic, AI boundaries and testing.
Technologies
React, TypeScript, Fastify, Zod, TMDB, OpenAI API, Vitest
01
The decision problem
The product narrows a high-choice catalogue using mood, runtime, content type, exclusions and Danish provider availability. The goal is a confident decision, not a longer recommendation feed.
Human control
Users can inspect and correct interpreted preferences before recommendations run.
Real catalogue evidence
TMDB metadata and provider availability remain the source of truth for every candidate.
02
Product flow and refinement
Free text becomes an editable preference summary. Deterministic filters create a valid shortlist, and users can reject, replace, request similar titles, edit preferences or restart.
03
Architecture and AI boundaries
The React client, Fastify API and shared Zod schemas keep interface state, server secrets and validation separate. AI extracts preferences and may rerank only an allowlisted shortlist; it cannot invent titles or metadata.
- Viewer request
Bounded text and explicit controls describe the current need.
- Preference extraction
Structured AI output is validated and remains editable.
- Catalogue and filters
TMDB data passes hard filters and inspectable scoring.
- Optional reranking
AI can reorder only validated candidate IDs with grounded evidence.
- Explainable result
The interface shows match reasons, compromises and provenance.
- Deterministic fallback
Recommendations still work honestly when AI is unavailable.
04
Testing, limitations and reflection
The local project includes schema, API and React tests plus browser QA from 360 to 1440 pixels. Production deployment and external usability testing have not yet been completed.
Status: Automated and responsive QA completed · External usability testing is planned.
Planned test details
Objective and participants
Test with 3–5 streaming viewers whether the request, preference correction and recommendation refinement flow remains understandable.
Tasks
- Describe a viewing need and interpret the extracted preferences.
- Correct one preference before requesting recommendations.
- Reject one result and request a replacement.
- Explain why a recommendation matches and where it compromises.
Observe and measure
Record completion, critical errors, assistance, hesitation, wrong turns and confidence in the explanations.
Pending evidence
Usability findings and resulting product changes remain pending until external sessions are completed.
What worked well
Deterministic validation keeps the product useful and explainable even when AI is unavailable.
Biggest challenge
Balancing a simple decision flow with provider, runtime, genre and AI failure states.
What I would improve
Deploy the edge architecture, add observability and test the complete flow with real viewers.
StudyMate AI
AI study prototype · Product and technical UX