Cultivar

Inspiration Independent specialty coffee roasters live and die by their seasonal bets. Every quarter, roasters invest scarce budgets and inventory into new single-origin drops, limited-edition blends, packaging runs, and local brand collaborations.

Unlike multinational chains with dedicated consumer insights teams, independent roasters rely almost entirely on gut instinct, social media noise, or generic trend reports. The challenge isn’t a lack of creative ideas—it’s finding credible, evidence-backed directions before committing capital.

We built Cultivar to level the playing field: transforming a roaster’s existing coffee lineup into cultural taste intelligence that uncovers authentic brand affinities, audience overlaps, and seasonal concepts with zero guesswork.

What it does Cultivar is a cultural intelligence platform that helps specialty roasters make informed seasonal product and collaboration decisions:

Catalog Ingestion: Roasters drag and drop their real product CSV or load a curated specialty lineup with a single click. Human-in-the-Loop Entity Matching: Cultivar maps catalog items against Qloo’s global taste database. Instead of trusting an opaque automated guess, roasters review and confirm entity matches with confidence scores. Deterministic Cultural Discovery: For confirmed catalog items, Cultivar queries Qloo to uncover real-world audience correlations across adjacent cultural domains (food & beverage, lifestyle, fashion, music, and outdoor goods). Transparent, Evidence-Backed Scoring: Every recommendation provides an inspectable signal breakdown—Taste Affinity %, Audience Overlap %, and Seasonal Relevance. Cultivar doesn't claim to predict sales; it reveals real cultural associations to investigate. Interactive Cultural Graph: An interactive network graph visually maps the relationships connecting catalog beans, shared cultural attributes, and recommended partner brands. Side-by-Side Baseline Comparison: A dedicated audit screen evaluates Cultivar’s grounded intelligence against a generic ungrounded LLM, proving how cultural data prevents generic clichés. How we built it Frontend & Architecture: Next.js 15 (App Router) with React, TypeScript, and Tailwind CSS. Cultural Taste Intelligence: The Qloo API powers cultural entity search and multi-domain affinity insights. Structured Synthesis Engine: Google Gemini 2.5 Flash (via @google/genai) synthesizes narrative collaboration angles, seasonal activation concepts, and tasting notes constrained to strict, deterministic JSON schemas validated with Zod. Deterministic Scoring Engine: Proprietary scoring algorithms in TypeScript that compute weighted affinities and audience overlap metrics directly from raw data—ensuring the LLM never fabricates metrics. Data Visualization: A custom, responsive SVG-based Cultural Graph showing dynamic node relationships, directional affinities, and contextual metadata overlays. Audit & Export: Local state caching and structured CSV/JSON export tools for roaster decision documentation. Challenges we ran into Specialty Coffee Entity Ambiguity: Coffee blend names are notoriously poetic (e.g., “Hair Bender”, “Holler Mountain”), often triggering false positives against songs or books in cultural databases. Rather than hiding this behind black-box logic, we designed a human-in-the-loop entity resolution stage (/resolve) where roasters verify or adjust candidate matches before generating insights. Separating Deterministic Data from Generative Copy: We wanted to ensure the AI never hallucinated metrics or inflated confidence numbers. We enforced strict architectural separation: raw affinities and scores are calculated mathematically from Qloo signals, and Gemini is strictly used to provide editorial context, tasting notes, and activation narratives. Real-World CSV Formatting: Real roasters use varied tools (Excel, Google Sheets, macOS Numbers, Notepad). We engineered robust catalog normalization that handles byte order marks (BOM), Windows CRLF line endings, case-insensitive headers, and accidental .csv.txt extensions. Accomplishments that we're proud of Inspectability Over "Magic": We avoided building another generic chat interface. Every recommendation card opens an evidence drawer showing exact signal sources, audience overlap figures, and data provenance. Side-by-Side Baseline Evaluation: We built a live comparison view that pits Cultivar directly against a raw LLM prompt, visibly demonstrating the difference between grounded cultural signals and generic AI fluff. Frictionless Real Roaster Experience: Roasters can test the platform instantly with pre-loaded real-world specialty coffee profiles or bring their own real-world inventory. Production-Grade Type Safety: Full end-to-end TypeScript validation from API responses through Zod schemas to client-side UI states. What we learned Small teams want context, not fake certainty: Independent business owners distrust platforms that promise “guaranteed viral hits.” Presenting signals as “evidence to investigate” rather than “sales predictions” creates immediate trust and genuine utility. Cultural taste data bridges the creative gap: Raw sales data only tells businesses what customers already bought in the past; cultural graph correlations reveal what those same customers care about outside the café doors. What's next for Cultivar Direct POS / E-Commerce Sync: One-click integrations with Shopify, Square, and Toast to auto-sync seasonal inventory and historical bean releases. Closed-Loop Feedback System: Allowing roasters to log the outcome of tested seasonal drops, building a private local memory dataset that refines future recommendations. Cross-Industry Expansion: Adapting Cultivar’s cultural intelligence pipeline for craft breweries, indie chocolatiers, and micro-distilleries facing identical seasonal product challenges.

Built With

• typescript