TasteDrop // The Cultural Cart Agent

Don't sell shoe grease to someone buying a rock-and-roll jacket. Autonomous cultural merchandising powered by Qloo.

Comment

Inspiration

Every e-commerce algorithm today behaves like a spreadsheet with a memory leak.

When someone decides to invest $650 into an authentic Schott NYC 618 leather motorcycle jacket, what does traditional cart software recommend at checkout?

  • A $12 plastic bottle of Bickmore leather cleaner.
  • An $18 pack of wooden coat hangers.

Technically, nothing is broken. It’s what people often co-purchase to maintain cowhide. But it completely misses the cultural context.

The person buying a Schott 618 isn't buying leather maintenance; they are buying into 1970s CBGB counterculture, rock-and-roll heritage, and rebellion. Recommending shoe grease at the highest-intent emotional peak of that purchase degrades the brand and leaves real cart value on the table.

I built TasteDrop around a single question: What if the checkout cart had taste? What if autonomous merchandising engines understood that a leather jacket lives in the same cultural universe as seminal proto-punk vinyl and dark amber fragrances, rather than cleaning rags?

What TasteDrop Does

TasteDrop is an autonomous merchandising agent that replaces generic maintenance commodities with cross-domain cultural companions.

Instead of showing shoppers utility items they don't care about, TasteDrop taps into Qloo’s cultural intelligence graph (250M+ entities) to understand the subcultural context of the primary item in the cart, discovers authentic cultural affinities across music, fashion, and fragrance, maps them against merchant catalog inventory, and assembles a unified co-curated bundle.

When you click "Run Qloo Taste Scan" on the Schott NYC jacket:

  1. The generic maintenance items are evicted from the recommendation slot.
  2. The agent pulls live cultural correlations from Qloo (anchored on proto-punk pioneers like The Stooges, MC5, and The Sonics).
  3. It recommends The Stooges: 'Raw Power' Audiophile Vinyl ($28.00) as a companion piece, followed by an artisanal post-purchase discovery (Maison Margiela Jazz Club).
  4. In the demonstrated scenario, the conventional basket contains $30 of utility accessories, while TasteDrop identifies a $28 cultural companion: The Stooges' Raw Power vinyl. The point is not a claimed revenue uplift; it is the ability to discover a culturally relevant cross-category companion instead of defaulting to generic maintenance accessories.

How I Built It: The Two-Layer Architecture

I designed TasteDrop with a strict architectural boundary so that AI is doing heavy cognitive reasoning, not hallucinating inventory.

Layer 1: Qloo Cultural Perception Engine

TasteDrop makes authenticated outbound HTTPS calls directly to Qloo's live API (https://hackathon.api.qloo.com/v2/insights). When an anchor item is scanned, TasteDrop queries Qloo’s entity taste graph, retrieving real cross-domain affinity correlations:

  • MC5: 0.9714 affinity
  • The Sonics: 0.9546 affinity
  • Iggy Pop: 0.9543 affinity

Note: These scores represent original Qloo-reported cultural affinity correlations, preserved through the matching and policy pipeline, rather than product-level purchase probabilities.

Layer 2: The 3-Tier Cultural Matcher & Deterministic Policy Solver

Raw Qloo affinities are cultural entities, not merchant SKUs. TasteDrop bridges this gap through a 3-tier relationship registry:

  • Tier 1 (Direct Creator Match): Identifies direct creator overlap.
  • Tier 2 (Cultural Concept Expansion): Traverses subculture taxonomy (e.g. mapping The Sonics $\to$ garage-rock $\to$ vintage apparel and vinyl) with auditable provenance.
  • Tier 3 (Metadata Token Alignment): Aligns inventory tokens with subculture aesthetics.

Once matches are found, my Deterministic Policy Engine enforces strict real-world merchant guardrails:

  • Hard Spend Ceiling: Total companion items cannot exceed $150.00.
  • Relevance Floor: Only entities with $\ge 80\%$ cultural correlation qualify.
  • Category Diversity: At most 1 companion per category (no duplicate clutter).
  • Autonomous Abstention: If no merchant catalog item satisfies the cultural matching and policy requirements, TasteDrop abstains and returns zero recommendations. The agent knows when not to sell.

Challenges I Overcame

  1. Bridging the Concept-to-SKU Gap: Qloo knows that people who love The Stooges also love MC5, but an e-commerce catalog sells jackets and vinyl records. I built a principled 3-tier taxonomy layer that maps cultural correlations to real merchant inventory while preserving the original Qloo-reported cultural affinity scores.
  2. Deterministic Restraint vs. AI Hallucination: LLMs love to recommend anything if you ask nicely. I enforced strict algorithmic policies: if an item is $360 (exceeding budget), it is rejected with an auditable reason. If affinity is below 80%, it is killed.
  3. Execution Latency: Observed end-to-end execution in testing ranged approximately from 60–250ms, depending on the live Qloo upstream response and network conditions.

What I Learned

I learned that cultural intelligence is the missing half of modern commerce.

Every e-commerce store has spent 15 years optimizing collaborative filtering ("people who bought this also bought this"), which inevitably degenerates into utility commodities (batteries, cables, cleaning spray). Qloo's taste graph enables a completely different paradigm: semantic cultural resonance. It turns transaction flows into editorial moments.

Forensic Verification & The "Glass Back Door"

I believe hackathon submissions should be auditable, reproducible, and verifiable by technical judges:

  • Live Production URL: Deployed and running live on Vercel at https://tastedrop-433o.vercel.app with live Qloo API key configuration.
  • 13/13 Green Unit & Regression Tests: Zero-dependency automated test runner (npm test) that proves adversarial inversion (poison inputs, null payloads, oversized numbers), parallel concurrency safety, and controlled causality experiments (verifying that Johann Sebastian Bach correctly forces 100% rejection against leather jackets).
  • Decision Lab HUD (Telemetry Drawer): Built directly into the frontend. Judges can click "⚡ Inspect Taste Graph" to see live latency in milliseconds, real-time candidate acceptance/rejection trails, relationship provenance, and copyable cURL traces to test the API directly from their terminal.

Built With

  • autonomous-agents
  • e-commerce
  • framer-motion
  • next.js
  • qloo
  • react
  • rest-api
  • tailwind-css
  • typescript
Share this project:

Updates

Submission history