I built a ranking system over about 200,000 expert profiles. It combined semantic and exact-term search with domain signals, reranking, and diversification. A trace showed why each result ranked, so the team could challenge the ranking and change it.
Domain experts could change complicated business rules and run real cases without writing code. When one failed, we separated a code bug from a missing or mistaken rule. That distinction became part of Sketch-CE.
The system combined 360-degree feedback, a skills rubric, and past reviews. It drafted a performance review and career-growth plan for someone to review and revise.
The system compared project briefs with available people, prior assignments, expertise, and performance evidence. It suggested a team for someone to accept, change, or reject.
For some work, a known case gives us something concrete to build and argue about.
Choose one with a known outcome. If no one can say what should have happened, we have more discovery to do.
Who does what? What do they look at? Where do the rules live? What happens when the answer is wrong?
Use code when we can state the rule. Use a model when the system has to interpret language or compare messy cases. Put the answer in front of the person who knows the job.
A bad result can come from code, a rule, the data, or our understanding of the job. Fix the layer that failed and run the case again.
The person who knows the work needs to try early versions. Otherwise we can make a polished system that misses the job.
Some projects are open source because I wanted to make them and can share them.
Coding agents can edit many files without knowing what depends on what. RNA builds a local graph from code, language-server data, git history, and repo artifacts. It lets an agent ask what a change touches, find code by meaning, or connect a commit to the business reason behind it.
It runs as one local binary and exposes the graph over MCP.
A coding agent can fix one example without capturing the rule that made it fail. Sketch-CE keeps an implementation repair separate from a business-rule change: people approve rule changes, while the agent can repair an implementation error.
The skills I use to decide what should change, compare approaches, do the work, check it, and keep what we learned when a plan fails.
A Rust bridge connecting Roon, LMS, UPnP, and HQPlayer to a physical knob, web UI, Apple clients, and MCP. I use it every day. It is also fun.
A weekend interview exercise that extracted fields from handwritten intake forms, sent uncertain fields to a reviewer, preserved corrections, and retrieved related cases.
We can work out whether I can help.