ReadNext

Qloo's taste graph and Gemini safety gates bridge student video game and TV passions into safe, captivating books.

Comment

💡 Inspiration

Every educator and librarian knows the reluctant reader dilemma: students who declare they "hate reading" will happily spend 20+ hours a week immersed in deep storytelling through video games (Minecraft, Roblox, The Legend of Zelda) or animated television series (Avatar: The Last Airbender, Pokémon).

Traditional book recommendation engines fail because they are trapped within literature silos: they ask for favorite authors, genres, or literary themes. When teachers turn to generic Large Language Models, they receive flat, stereotypical suggestions: ask an LLM for books for a Minecraft fan, and it simply recommends official merchandise like Minecraft: The Island.

We built ReadNext to break media silos by projecting a student's entertainment passions across domains into genuine, captivating literature powered by Qloo's cultural taste graph.

🚀 What It Does

ReadNext is an agentic recommendation platform tailored for K–12 teachers and school librarians:

  1. Intuitive Student Profile: The educator selects an age band (Primary 6–10, Middle School 11–14, or Teens 14–18) and picks 1–4 favorite video games or TV shows.
  2. Cross-Domain Taste Projection: The ReadNext agent queries Qloo's taste graph across cultural boundaries ($\text{Video Game / TV Show} \longrightarrow \text{Literature}$) to identify books sharing underlying narrative DNA, exploration mechanics, and emotional agency.
  3. Empirical Explainability: Every recommendation displays mathematical signal affinity scores derived directly from Qloo vector explainability: $$\text{Affinity Score} = S(e_{\text{input}}, e_{\text{book}}) \in [0, 1]$$ The agent grounds its explanations in these empirical weights rather than inventing hallucinations.
  4. Dual-Layer Classroom Safety Gate:
    • Deterministic Code Gate: Instantly filters explicit/adult media tags and enforces reading-effort constraints (e.g., dropping books $> 250$ pages when "Short Reads" is enabled for reluctant readers).
    • Gemini Flash Safety Gate: Batch-evaluates descriptions for age appropriateness, screens potential sensitive triggers, and generates concise 25-word teacher insight notes.
  5. Interactive "Compare with LLM-Only" Mode: A side-by-side view demonstrates why Qloo's taste graph outperforms standard LLM prompting: generic LLMs rely on literal keywords and franchise spin-offs, whereas Qloo uncovers deep literary connections (e.g., mapping Minecraft to Gary Paulsen's Hatchet and Peter Brown's The Wild Robot).

🛠️ How We Built It

ReadNext is built as a full-stack, secure TypeScript application:

  • Qloo Hackathon API: All upstream calls query https://hackathon.api.qloo.com. We utilize explicit type-scoped searches (urn:entity:videogame, urn:entity:tv_show) to prevent popular gaming titles from collapsing into merchandise brands, and query /v2/insights with fine-grained genre tag filtering (childrens, juvenile, middle_grade, teen).
  • Autonomous Agentic Pipeline: Bound by strict operational limits (maximum 6 tool calls and 3 LLM calls per request). If initial safety filters yield fewer than 3 books, the agent automatically triggers a tag-relaxation retry before finalizing recommendations.
  • Frontend & UX: Built with React 19, Vite, and Tailwind CSS. Features one-click judge presets for instant evaluation in under 30 seconds, live agent execution traces, and zero-login access.
  • Enterprise Security & Compliance: Strict server-side secret isolation (QLOO_API_KEY and GEMINI_API_KEY never leak to the client), per-IP rate limiting (10 req/min), in-memory LRU caching with zero disk/repo storage of Qloo data, and an atomic daily quota tracker to protect API budgets.

🧗 Challenges We Ran Into

  • Entity Disambiguation & UUID Formats: Untyped queries often returned TV show spin-offs or brand entities. Enforcing strict URN types and building automatic UUID format resolvers ensured zero malformed upstream requests.
  • Strict School-Safe Content Alignment: Demographic age signals can occasionally drift into adult genres. We bypassed generic demographic filters in favor of fine-grained literary genre tags combined with an LLM evaluator acting as a certified school librarian.
  • High-Availability & Latency Optimization: To ensure seamless judge evaluations during API traffic spikes, we engineered an in-memory LRU cache and automatic model fallbacks between Gemini Flash models with graceful degradation.

🏆 Accomplishments That We're Proud Of

  • Genuinely Agentic Architecture: Transparent live execution timeline showcasing tool invocations, tag relaxations, and safety audits.
  • Side-by-Side Proof: The "Compare with LLM-only" toggle clearly and objectively proves Qloo's unique competitive advantage to any evaluator.
  • 100% Verified Smoke Test Suite: 16/16 automated unit and pipeline assertions covering tag taxonomy, cache eviction, code-level safety filters, and quota guards.

📚 What We Learned

Qloo's cultural taste graph demonstrates that cultural affinity is not confined to individual media formats. Children don't love Minecraft merely because of "pixels" or "blocks"—they love it for the autonomy of survival, creative construction, and wilderness exploration. By mapping those implicit signals across media lines into books, we can turn screen time into reading time.

Built With

Share this project:

Updates

Submission history