SUTO OSNOVA — agentic cultural intelligence

Agents, but with taste: an agent that stops guessing what people love and measures it on Qloo's taste graph instead.

Comment

Inspiration

The pitch of this hackathon is real: agents can reason and plan, but ask them what people actually love and they guess. I wanted an agent that ANSWERS with evidence — cultural taste measured on Qloo's taste graph instead of model confidence. And as the challenge itself puts it: if the submission would work the same without Qloo, you're building the wrong thing — so the graph had to be the product, not a decoration.

What it does

SUTO OSNOVA turns a plain cultural anchor ("David Lynch Noir", "Glastonbury Festival") into a decision-ready dossier in one request:

  • Worldbuilder — mise-en-scene scene dossier: top affinities, sonic, literary, material, gastronomic and visual layers;
  • Experiential — live-event activation: resonance index with weighted evidence, curation, audience insights, food zones;
  • Taste antithesis — the cultural antipode with divergence percentages per domain.

Exports as JSON (schema suto-osnova.dossier/v1), Markdown, or print-ready HTML/PDF. No accounts, no storage: the page stores nothing and tracks nobody. Live demo: https://suto-osnova.netlify.app

How I built it

Vanilla ES6+ front end (zero frameworks) over a Node.js 22 serverless gateway on one origin. Every taste fact comes from four Qloo endpoints — /search, /v2/insights, /v2/tags, /v2/audiences — behind one hardened transport: X-Api-Key, sliding-window rate limit, bounded retries, two-tier cache, single-flight, privacy guard. The engine math (adaptive thresholds, pop-cliche filter, a uniqueness coefficient) is pure and deterministic, so the whole cycle also runs offline against recorded fixtures. The optional LLM narrative is bound to the measured numbers by a validator. The MCP surface (18 tools, qloo_* namespace) is validated against the official Qloo harness (npm run mcp:verify), and 1,200+ tests run on every push. Everything was built inside the submission period — first commit 2026-10-15, full history public.

Challenges I ran into

Qloo rejects some entities as signals — even against their own domain. Rather than paper over that, the resolver PROBES the signal and swaps in a comparably-scored alternative, or fails closed with a typed refusal (no_usable_signal): a wrong answer is never worth a fabricated one. Balancing the uniqueness filter — measuring how un-crowded a taste is without burying genuinely popular picks — took real iteration.

Accomplishments that I'm proud of

  • The API Inspector: the demo shows its own wiring — the client call log, live Qloo transport state (rate limiter, cache counters) and the backend's provenance ledger, verbatim. Judges can audit exactly what Qloo was asked.
  • A fail-closed agent: when Qloo cannot answer, it refuses instead of inventing taste.
  • A permanent uptime sentinel guarding the demo through the judging window, and pixel-locked brand assets (the cover is palette-pure down to the pixel).

What I learned

How far cross-domain affinity math takes you: one anchor seeds five domains at once, and measured divergence is a far better editorial signal than a model's confidence score.

What's next for SUTO OSNOVA — agentic cultural intelligence

The measured layer already returns dining recommendations from the graph — the hackathon's lateral directions (room moods, health-aware dining, community journeys) are parameter sets on this pipeline, not new products. Next: more presets for events and studios, and deeper MCP integration for any AI agent.

Built With

Share this project:

Updates

Submission history