TasteForge Agent

An evidence-first creative direction agent that turns cultural seeds into Qloo-grounded cross-domain campaign strategy with visible tool traces.

Comment

Inspiration

Generic LLMs can write plausible creative briefs, but plausibility is not cultural evidence. A creator or small brand still has to guess whether a proposed film, artist, brand, or place actually belongs near the audience taste they are targeting.

TasteForge turns that guess into an inspectable Qloo workflow.

What it does

TasteForge takes 2–4 cultural seeds plus a campaign objective.

  • Qloo describe resolves every seed that is allowed to influence synthesis.
  • entity_tags extracts cultural concepts shared by the resolved seeds.
  • recommend finds adjacent brands, movies, artists, and places.
  • With four resolved seeds, compare_audiences compares seeds 1–2 against seeds 3–4 as two cultural poles.
  • A Creative Tension Map uses only named shared affinities and differentiators returned by that comparison.
  • The final campaign brief is synthesized only from returned Qloo evidence.
  • Every Qloo operation remains visible with status, correlation ID, warnings, duration, retry count, and provenance.

If a workflow fails or returns no evidence, TasteForge leaves that domain unresolved instead of replacing it with a model guess.

How I built it

TasteForge is a Node.js application using the official @qloo/[email protected] workflow boundary, pinned to workflow contract 1.0.0 and result schema 1.0-preview.1.

The Qloo key is server-side only. The browser sends bounded seed/objective/market data to POST /api/forge; the server executes official Qloo workflows, normalizes versioned result envelopes, and returns separated evidence, synthesis, and trace layers.

The default judge flow sends explicit entity types with each cultural seed. This removes avoidable name ambiguity while preserving the fail-closed rule: only entities successfully resolved by Qloo can influence synthesis.

The project includes GitHub Actions CI, deterministic integration tests, typed-seed regression coverage, credential-exclusion tests, runtime-contract verification, a one-shot retry for transient Qloo rate limits, and a strict submission-preflight gate. The public demo is deployed on Vercel.

Live Qloo proof

On October 6, 2026, the production deployment completed the real judge flow against the Qloo Hackathon API:

  • 4/4 typed seeds resolved: Jil Sander, Brian Eno, Lost in Translation, Blade Runner.
  • 10 official Qloo workflow steps recorded.
  • entity_tags returned 8 cultural tags after one transparent retry of a transient 429.
  • brand / movie / artist / place recommendation workflows returned 6 results each.
  • compare_audiences succeeded.
  • the Creative Tension Map rendered in the public UI.
  • production health reports qloo_configured=true.
  • npm run submission:strict returns ready=true.

The retry is bounded: one transient Qloo rate limit is retried once; otherwise the workflow fails closed rather than inventing evidence.

Challenges I ran into

The hardest part was making Qloo the actual source of truth instead of a decorative enrichment call. I also had to keep partial failures honest, avoid treating affinity values from separate Qloo calls as one comparable scale, and design the four-seed Creative Tension Map so it preserves contrast rather than flattening everything into a generic moodboard.

Another challenge was live reliability. The hackathon endpoint can return a transient rate limit, so TasteForge records the retry count and retries exactly once before exposing the failure. Credential hygiene remained non-negotiable: the Qloo key never reaches the browser, repository, screenshot, deployment receipt, or public live artifact.

Accomplishments that I'm proud of

  • 17/17 deterministic tests passing.
  • Official Qloo runtime contract verified.
  • Real production Qloo run completed with 4/4 resolved seeds.
  • Public Vercel deployment and health endpoint live.
  • Creative Tension Map grounded in one official audience-comparison operation.
  • Missing evidence remains explicit instead of being hallucinated.
  • Git history and public evidence artifacts are credential-clean.
  • Strict judge gate returns ready=true.

Evidence

What I learned

Agentic cultural intelligence becomes much more useful when provenance is part of the product. The important output is not just a recommendation; it is a recommendation whose cultural evidence can be inspected, challenged, and refined.

What's next

Next I would add reusable campaign workspaces, richer Qloo comparison visualizations, more typed seed domains, and exportable creative-direction packages while keeping the same evidence-first and fail-closed contract.

Built With

  • agentic-ai
  • ai-agents
  • cultural-intelligence
  • javascript
  • node.js
  • provenance
  • qloo
  • qloo-harness
  • vercel
Share this project:

Updates

Submission history