CoastLens · how it was made
AI did the evidence archaeology. It is kept out of the physics.
Ireland's coastal evidence is real but scattered across agencies, formats, resolutions and decades. Three people could not have assembled it by hand in a month. AI could, with a person checking every step. And then the important part: the model is not in the result.
Where AI is
Agentic research. Agents ran across the public record and retained what they found, with source, licence and date on every item. The retained audit, counted conservatively:
That produced the source register, the 87-storm national index, the vertical datum chain, and the finding that only 98.3 km of Wexford's 245.3 km coast has a public elevation baseline.
Agentic evidence work, live. On 11 and 12 September, agents found that the national coastline service was capping at 2,000 records and paged it (1,047 features became 3,845); built the census and planning layers; found the OPW flood extents, read the licence, and refused them; and audited the commercial figures against sources, correcting one. Every run is on the Agents page with its timestamp. A person approved each one.
The build. The engine, the GIS and this site were written with AI assistance in under two weeks by three people working evenings, with unit tests on the physics at every regime boundary. Where the AI proposed a scientific choice, a person made it.
Where AI is kept out
The result. Stockdon et al. (2006) and Sallenger (2000) are published equations. A language model never touches a regime label. It sits behind a policy gate that rejects unknown source IDs, unsupported numerical claims, missing caveats, changed regime labels and prohibited language about engineering works.
The gaps. When an input is missing the engine returns NO_DATA and the screen shows it. Storm Kathleen's evidence-grade result at Sea View is NO_DATA today, and we did not overrule it to get a better demo. What the demo shows is a Kathleen-informed prototype scenario, labelled as such on the display.
The decision. Agents prepare. The engine calculates. A named council officer approves what leaves. Nothing on this site has been sent to an external authority.
Why it could not be done as well without AI
Not because the physics needs it; the physics is thirty years old and deterministic. Because the evidence does. Finding, paging, reprojecting, licensing and registering a county's coastal record is analyst time, and analyst time is the constraint on every coastal team in the country. AI turned that from a month of three people into a weekend with one person checking. That is the service: a maintained evidence base a council could never afford to keep by hand, feeding an engine whose answer a council can defend.
Tools, and how each was used
Manus (Challenge resource partner): agentic research. Twelve bounded briefs, each closing a named decision or interface field, each registered before launch and reviewed on return. It produced the source register, the storm index, the datum chain and the coverage finding. It was never asked to decide a regime or invent a saving.
Claude Code: the build and the evidence agents. Engine, GIS, layers, this site. Supervised sessions with a written scope; the refusals and corrections are kept in the record and shown on the Agents page.
Kept out of the model on purpose: the physics (Stockdon 2006, Sallenger 2000, unit tested), every regime label, and every number that reaches a council. Responsible AI here is not a policy page; it is the policy gate the system runs behind.
MapLibre for the GIS, Vercel for hosting, open data throughout. Every dataset and its licence is listed on the Q&A page.