Case 03 / Government-affairs strategy · Data collection
I built a tool to collect strategic judgment from experienced government-affairs practitioners when no suitable public dataset existed. It turns their knowledge into structured examples for model evaluation and later training, supporting the strategy system in case 02.
01 / The problem
The strategy system needed examples of how to make good decisions in government affairs. That knowledge lived with a small group of experienced practitioners. There was no suitable public dataset of vetted strategies in the required format; bills, filings, and news could supply context, but they could not supply those practitioners’ judgment.
I built a collection tool to draw that judgment out of specialists and turn it into consistent, structured examples a model could use later. The specialists were busy senior practitioners, so the tool had to guide them through a rigorous contribution without requiring them to think like data engineers.
02 / My decisions
The instrument guides specialist practitioners through the strategy in a shared format and vocabulary, capturing decisions and reasoning as structured examples a model could later use. This avoids reconstructing structure from inconsistent write-ups afterward. The cost is more friction during contribution.
Public sources can inform an expert’s work. They do not become ground-truth labels just because they are available. The labeled example is a practitioner’s vetted strategy.
I prioritized a trusted evaluation set before fine-tuning. Its first purpose was to measure the grounded system. Training was deferred until the dataset was sufficiently large and trusted.
I designed a working authoring instrument, an authoring kit, and standards for a good example. The collection workflow was a product experience in its own right, not just a storage mechanism.
03 / Delivery
I delivered the authoring kit, working instrument, and example-quality standards. Data collection was still underway at handoff. Fine-tuning was intentionally deferred, so I do not claim a trained-model result or measured quality improvement.
Based on my anonymized AI portfolio. The delivered output was the collection capability; the dataset and downstream model work were not complete at handoff.
Keep exploring