Dark Sky

The client's prior deliverable had been produced with generative image models. It looked convincing and was wrong: a transmission line had moved from its surveyed position, and a pole inventory was miscounted. For work that feeds engineering and siting decisions, plausible-looking output is worse than no output — it carries the authority of a survey without the accountability of one. They needed overlays and viability studies their own customers could stake a decision on.

• Built a deterministic CAD-to-overlay pipeline: survey CAD and classified point-cloud/GIS data placed on a georeferenced orthomosaic, where every point, line, and polygon lands where the data says it lands
• Enforced the guarantee in code — a static analysis guard and a runtime import guard that fail on every test run if generative model code enters the overlay path at all
• Constrained the study writer so it can smooth prose but cannot introduce a single number the statistics layer didn't compute, checked by a digit-level diff on every text surface
• Exposed the platform three ways — Python library, console tools, and an agent-callable server — so it fits existing workflows instead of replacing them
• Delivered a written scope-and-limits document stating precisely what the no-AI guarantee does and does not cover
• Transferred ownership at handover, with the client holding the code outright

The client can now put a number in front of their own customer and say where it came from. The determinism claim isn't a promise in a sales deck — it's a test that fails the build, which means it survives staff turnover and future changes. Ownership transferred on day one of the agreement, so the platform is theirs, not rented.