AI Builder, Devekko.dev / Agentic ecommerce
California Wine
Connect a wine preference to the winery behind it.
California wine discovery built around producers, regions, and preferences, with an agentic consumer-to-winery commerce concept.
Demo and research snapshot: October 29, 2026. Proposed scope is labelled below.
The product question
Can better wine information help a consumer choose a producer and make a useful inquiry?
- Describe a preference
- Explore wines
- Visit the winery
Product intent
A consumer choosing California wine may know a taste, an occasion, or a budget before they know a producer. This project explores how a structured wine directory and an AI-assisted conversation can connect that preference to a winery.
The commercial hypothesis is a better handoff: a consumer understands the wine and the producer receives a more useful visit or inquiry. The winery remains responsible for its offer and purchase experience.
What the demo shows
The public site provides California wine and producer discovery, collections, regional information, and reference prices tied to sources. Visitors can explore wines by characteristics such as grape, region, style, and price, then follow winery purchase links.
Purchases take place at winery sites. Reference prices are dated research, not a promise of current stock or a checkout quote. The directory is independent; inclusion does not establish a winery partnership. Consumer-to-winery agentic commerce is the broader product direction.
Product specification
Primary users: consumers looking for California wines, winery teams reviewing inquiries, and editors maintaining product information.
Core workflow: express a preference, review suitable wines with the reasons and sources behind each suggestion, build a shortlist, and choose a winery handoff. A future inquiry flow should show the exact message and recipient before sending.
The proposed next release should support:
- A structured catalogue linking wines, producers, regions, source URLs, and checked dates.
- Preference capture that allows a shopper to correct an assistant's assumptions.
- Recommendations grounded in catalogue facts, with an explicit response when evidence is missing.
- Reviewed winery onboarding and correction requests.
- A consent-based inquiry handoff with clear delivery status and operator review where needed.
Acceptance means a recommendation can be traced to catalogue facts, stale prices are labelled, and the consumer can reach the winery without mistaking the directory for the seller. Any future checkout must separately address merchant responsibilities, age requirements, and shipping eligibility before launch.
Technology
The current application uses TypeScript, Hono, Cloudflare Workers, D1, and a static frontend. A separate Python intelligence service supports the project architecture; container hosting and Postgres options are documented independently of the public discovery experience.
Pydantic AI enrichment is designed to prepare metadata drafts from supplied producer, wine, and retailer pages. That path depends on crawler and model configuration and reviewed publication. It should not be described as an always-on system that changes prices autonomously.
The broader sprint toolkit includes Codex and Claude for building and review. PlanetScale Postgres, Prefect, Nemotron, and Brave API are options within the sprint's intelligence work, rather than a claim that every wine workflow uses them in production.
Research and development
The R&D plan tests the usefulness of the handoff:
- Compare catalogue filters with conversational preference capture on the same wine-selection tasks.
- Review extracted product facts against original winery pages, including vintage, price date, and conflicting descriptions.
- Test whether an explanation helps consumers choose or simply adds reading time.
- Ask winery operators which inquiry fields are sufficient to respond and which add unnecessary friction.
Agentic workflows should draft recommendations and route information within explicit permissions. Purchase commitments and outbound messages need a defined approval boundary. User and winery interviews are proposed research, not completed findings.
Business measures to test
Measure discovery-to-winery click-through, shortlist completion, qualified inquiry completion, and accepted winery referrals. Operational measures include source freshness, correction rate, review time per wine, and cost per useful recommendation.
Revenue attribution would require a confirmed winery relationship and an agreed measurement method. The sprint does not report sales uplift or completed multi-winery transactions.
Sources and next release
The public experience was reviewed on October 29, 2026: California Wine. The technical description also uses the project's implementation notes and content boundaries. Authenticated operator workflows were not exercised for this review.
The next decision is whether discovery quality and a reviewed winery inquiry create enough value to expand the agentic commerce flow.