The Pivot: Why I’m Changing Course

Through the Discovery process and a great coaching session with my Hack Your Summer mentor, I decided to adjust my project direction. My primary career goal right now is positioning myself for future job opportunities and interviews, and my mentor gave me a clear piece of advice: focus on agentic AI, tool orchestration, and evaluation harnesses.

While I still care about the water quality for brewing project, building out a true agentic pipeline for environmental reports requires a lot more domain research and user interviews. Pivoting back to my main open-source passion project, OpenBreweryDB, lets me iterate faster on open data while learning how to build an agentic system from scratch. Well, at least instructing AI to build one by scratch.

Maintaining an open database of 11,000+ breweries requires constant verification (or lack thereof lately, in our case). Thankfully, the community has stepped up, and I want to build better tools for both them and myself. Spot-checking whether a venue has closed, moved, or updated its details takes several minutes per brewery of manual tab-hopping, searching, and schema editing. Even more so when you have ADHD.

I’m building BreweryAgent: a dedicated, local-first research engine designed to automate this entire discovery and verification loop.

The Interface & Tech Stack

My initial thought was to build a public Discord bot right away, but open feeds carry risks of prompt injection and API token abuse. Instead, I’m building this as a local-first personal research tool, though I still plan to open-source it for other contributors.

To iterate quickly, I’m building the frontend in SvelteKit, a stack I know well that easily handles our state and UI requirements.

The MVP Canvas Summary

1. Project Vision

  • What it is: A local-first agentic research tool that automates the web search, page scraping, geocoding, and schema diff generation needed to update OpenBreweryDB records.

  • Target User: "Contributor Cody," a dataset contributor or open-source maintainer who wants accurate brewery data without wasting hours doing manual Google searches and hand-editing JSON files.

  • Core Value Prop: Accelerate dataset updates from 10 minutes down to seconds while ensuring high accuracy through strict budget and evaluation guardrails.

2. Core Feature Set

  • Custom ReAct Engine: A pure Python Thought → Action → Observation loop built from scratch to orchestrate multi-step research.

  • Tool Suite:

    • OpenBreweryDB API: Fetches existing records.

    • Web Search & NER Page Scraper: Searches news/blogs and extracts structured metadata (e.g., website, phone, status).

    • Geocoder: Validates latitude and longitude coordinates (likely using Geocodio).

  • Budget & Rate Governor: Hard execution caps (e.g., max 3 web searches, $0.02 total token spend per run) to prevent runaway costs or infinite loops.

  • Offline Evaluation Harness: A stand-alone test runner benchmarking the agent against 10 ground-truth news snippets to measure Schema Extraction Precision and Hallucination Rates.

  • SvelteKit UI: A clean local web dashboard to review research findings, inspect agent logs, and approve proposed dataset updates.

MVP Canvas (Google Doc)