Brew Guide

The problem

Brew parameter guides are anecdotal and inconsistent — different sources recommend wildly different temperatures, ratios, and grind sizes with no basis in collective data.

My role

Sole developer — designed and built the MCP server, recommendation engine, database schema, and all tooling end-to-end.

Highlights

• Engineered a deterministic recommendation engine that calculates confidence tiers (High/Medium/Low) based on community data volume rather than LLM probability.
• Designed a 3-tier origin verification system (Exact, Fuzzy, Unknown) that programmatically degrades recommendation confidence to prevent 'hallucinated' consensus from imprecise data.
• Implemented a suite of 5 MCP tools over a public Streamable HTTP endpoint, enabling agentic coffee brewing via Claude Desktop, Cursor, and other MCP clients.
• Architected a dual-journey API flow (Query → Log → Compare) supported by a unified snake_case data model and Neon Postgres + Prisma ORM for strict type safety.
• Developed a 'Delta Analysis' engine that provides structured feedback on variance across 4 dimensions (temp, time, ratio, grind) relative to method-specific defaults.
• Refactored the entire persistence layer from local sql.js to a cloud-native Neon Postgres architecture while maintaining 100% TypeScript strict-mode compliance.
• Established a rigorous CI/CD pipeline with a 56-test Vitest suite and auto-deployment to Railway, ensuring zero-regression logic for complex scoring algorithms.
• Designed a 'Literature Automation' scaffold to ingest unstructured forum discussions, bridging the gap between community tribal knowledge and relational data structures.

Architecture

Brew Guide is a community-driven recommendation engine that aggregates coffee brewing experiments into deterministic weighted consensus parameters for specific origins and methods. The system is built on Node.js 24 and Hono 4, utilizing Neon Postgres with Prisma for persistence and Vitest for a rigorous test suite. A key architectural decision is the exposure of the engine via the Model Context Protocol using a streamable HTTP transport, allowing AI agents to query brewing data without relying on LLMs in the hot path. This setup ensures high-confidence recommendations categorized by data density tiers while maintaining an open-access origin policy for seamless integration with tools like Claude Desktop and Hermes.

Impact

Public MCP endpoint at brew-guide-production.up.railway.app/mcp — usable by any MCP-capable tool. Submitted as an entry to the Hermes Agent Challenge 2026.

Stack