In Prod

The technical deep-dive: JSON-RPC over stdio, FastMCP, asyncpg connection pools, the JSONB codec gotcha, tool descriptions as routing logic, and a four-tier test strategy. Everything glossed over in the methodology post.

How we built a custom MCP server for Shelf in one autonomous session — 7 tools, 43 tests, live competitive intelligence queryable from Claude Code. The spike branch pattern and what failed when we hit the real database.

We shipped VizStack as an open-source architecture visualizer. Then we wired Claude directly into the graph. It became an onboarding tool. We didn't plan that.

The real eval stack behind migrating a production LLM prompt on Shelf — a 32-check scoring rubric, a one-way runner, 54 adversarial payloads, Pydantic strict validation, and a hard rule about who can approve a prompt edit.

We used Langflow to sketch agentic patterns before hardcoding them in Python. We wired it to a local Postgres with production-level data. Here's what we learned and where we're still stuck.

What 250K context actually feels like in production. How we structure long-running prompt jobs on Shelf, what the industry is beginning to understand about this shift, and why we haven't really cracked it yet.

Why your app architecture is harder to hold in your head than it should be, and what we built to fix it — a route tree, an outlet chain view, and a cross-layer ecosystem map.

vizstack is a zero-install CLI that turns any Remix or Next.js codebase into an interactive dependency map. Point it at a directory and it produces a single HTML file showing your routes, database, AI providers, and external APIs as a scrollable graph — hover any node to see what connects to it.

Open interactive map →

How we built a production Shopify competitor intelligence app end to end — the stack, the pipeline, the prompting mistakes, and what the discipline actually gave us.

Shelf is a Shopify app that gives merchants a daily briefing on what their competitors are doing — prices, inventory signals, stock levels. It crawls rival storefronts with Playwright, passes the signals through a four-layer AI pipeline, and surfaces the output inside the merchant's Shopify admin.

What humans bring to AI systems — taste, judgment, and technical intuition.