Cool Vibes, Cold Numbers: Learn AI Coding for Sharper Stratification (Half-Day Workshop)Monday, February 22, 2027 - 8:00 AM - 12:00 NoonInstructors: Jonathan Beck, AAS, Assessor, Beckham County, Oklahoma Jake Parkinson, Recovering Assessor, Tyler Technologies, UtahAssessors are being told that artificial intelligence will change their work. Almost nobody is teaching them to drive it. This workshop does — four hours, hands on the keyboard, building working analytical tools from scratch, with no prior programming experience required or assumed.
The subject is AI-assisted development as it must be practiced inside an assessment office. That means guardrails before code: a source-of-truth hierarchy the model is not permitted to override — statute first, then standard, then office policy, then convenience — along with explicit scope limits and a definition of "done" written before the first line is generated. It means choosing an architecture on purpose rather than accepting whatever the assistant offers first and understanding why a single locally-run HTML file is almost always the right output target for a public office, while a web application is an entirely different conversation with your IT department.
And it means never handing over your data. Attendees learn to build from column names, not cells. You give the model your schema — field names, data types, what each field means — and it writes the tool. The tool runs on your machine. Your parcels go into the tool, never into the chat. Close the tab and nothing was retained, nothing was transmitted, and nothing entered a training corpus. For attendees in nondisclosure states, that is not a convenience; it is the only version of this work that is lawful.
Nobody will sit staring at a blank prompt box. Example prompts are supplied throughout — guardrail language to paste and adapt, build premises for each lab, debugging prompts for when a tool breaks, and the phrasing that gets a model to explain its own logic rather than defend it. They are a starting point rather than a script, and the compare sessions after each lab exist so attendees can trade their own improvements around the room.
All materials live in a shared Google Drive folder. Attendees receive access before the conference — the synthetic dataset in two sizes, the data dictionary, the guardrail checklist, the architecture decision sheet, and the prompt library — so setup problems get solved at home instead of in the first twenty minutes of class. The folder stays live during the session: attendees drop their own builds and working prompts into it as they go, which is how thirty parallel experiments become one shared library. It remains open afterward, along with everything the room produces together.
The session runs as three teach-build-compare cycles: a short instruction block, a build problem, then the room reconvenes to compare results and trade prompts. Lab 1 produces a data integrity test — year built against total area, price per square foot against a cohort of comparable homes, whatever inconsistency each attendee chooses to hunt. Lab 2 adds the stratification layer: segment the data, compute ratio statistics per segment, and surface cell counts beside every statistic so that thin segments announce themselves rather than hiding. Lab 3 turns that analysis into something presentable to a board, a commissioner, or a taxpayer at the counter.
Between Labs 1 and 2 sits the block most sessions skip: iteration, debugging, and verification. What to do when the tool breaks. How to recognize output that is plausible and wrong, which is far more dangerous than output that crashes. How to check work you did not personally write — because a value you cannot verify is a value you cannot defend, and no model takes that responsibility off your desk.
The problem all this builds toward is a real one. Aggregate ratio compliance can conceal significant sub-market inequity, because the statistics regulators require never ask where and never ask when. By the final lab, attendees have built a tool that surfaces those pockets — discovered firsthand, rather than watching a presenter find them on a slide.
Attendees leave with a working tool, the process that produced it, a shared library of prompts and builds from everyone in the room, and the judgment to know when to stop building.
Keywords: Artificial Intelligence & Machine Learning, Statistics & Ratio Studies
