Fly fishing conditions app covering nine rivers across Maryland, Pennsylvania, and New York. Every Sunday at 6am — and on demand via webhook — an n8n pipeline pulls USGS gauge data, regional weather, and scraped fly shop reports, then runs nine parallel Claude calls that synthesize each river's conditions in a guide's voice. Output commits to git and auto-deploys. Built because I fish these rivers, which makes for the most honest QA panel available: fishing buddies who text me when the report is wrong. The micro-app thesis, tested on something I actually care about.

Daily content monitoring loop running at 5am ET without manual intervention. Gemini handles high-volume triage scoring across monitored sources; Claude synthesizes only the flagged items. Output surfaces in a Google Sheet. Two-model architecture — separating triage from judgment keeps the synthesis pass focused and cost controlled.

Customer Success simulation environment built around a four-agent chain: Monitoring detects signals, Prioritization scores urgency, Research assembles record context, Responder drafts output, Handoff delivers to dashboard. Three SQLite databases serve as shared agent memory; nine MCP tools exposed for external access. Accounts with no signals produced silence — technically correct, but even knowing the system worked, the absence looked like failure. The trust problem is harder than the technical one.

n8n workflow routing identical prompts to multiple local Ollama models simultaneously. An LLM judge evaluates and scores each output against defined criteria. Results surface in a review UI with approve/reject controls, a notes field, and CSV export. Built to develop intuition for model behavioral differences before applying them in production patterns.

AI assistant deployed inside approved enterprise tooling to support analyst quality review. Structured persona, defined operational guidelines, asynchronous availability across distributed timezones. The constraints were the design — built to function within what the organization already permits, not around it. First production AI deployment.

Prompt, skill, and context file library with automated quality scoring, LLM-powered refinement, and version history. Built to solve the problem of distributing AI context inside an enterprise without requiring others to change how they work. During early testing, a hallucinated rewrite was plausible, structured, and well-worded — the system had no way to know that "improved" and "accurate" are not the same thing. That gap is what the tool is built to surface.