No app to build. Ask in your own words, and Eunice pulls the data, does the analysis, and hands back the finished map, report, or spreadsheet, with its sources.
Eunice works out what the question needs and does each step herself: 19 steps, 2 minutes 13 seconds.
She models the release forward two hours on the forecast wind at the time of the run. Each point of light is the center of one modeled puff, and every puff spreads as it travels.
Then she draws the screening area, everywhere those puffs could reach, drawn wide on purpose, and finds every school, hospital, childcare center, and nursing home on record inside it.
And she says what the numbers are, and what they are not.
Hypothetical release. Modeled screening estimate. No release rate was given, so Eunice assumed 10 kg/s and said so. Counted inside the screening area, which allows for each puff's full spread: facilities that could be reached, not confirmed exposures. Facility locations: HIFLD, dated records 2010 to 2024. Run October 20, 2026.
"A tank at 5600 Douglas St, New Orleans is venting formaldehyde. Which schools, hospitals, childcare centers, and nursing homes are in the path over the next two hours?"
Hypothetical release. Modeled screening estimate. No release rate was given, so Eunice assumed 10 kg/s and said so. Counted inside the screening area, which allows for each puff's full spread: facilities that could be reached, not confirmed exposures.
Facility locations: HIFLD, dated records 2010 to 2024. Run October 20, 2026.
The record of the same run, as it was kept.
Answers from your document library carry the exact source passages they came from, so an answer is something you can check rather than take on faith.
Sending email or changing a ticket asks first. Merging records across systems waits for an authenticated approver in an approvals ledger, with two approvers for higher tiers.
Every message, tool call, input, output, and cost is recorded without truncation, and any conversation can be replayed.
In the run's own words: "plume footprint" is the screening area above, and "particle rows" are the modeled puffs.
You have a question about your own data.
It becomes a ticket.
It waits in a queue.
An expert pulls the data.
The answer comes back in their terms. Often, it goes around again.
With Eunice you ask directly, and the answer comes back in your terms, with its sources. Your expert reviews it instead of fetching it.
Eunice does not add expertise to your organization. She stops rationing the expertise you already have.
A chemical release, asked in plain language. She screens the facilities in its path, then audits her own run, including what she could not do.
Uploads, documents, exports, and connected sources, each with its lineage and its history of use, grouped into kits.
A repeatable review written once in readable prose, with the steps it requires and the inputs it needs, kept in versions.
Every message, tool call, input, output, and cost, recorded without truncation and open to inspection.
Every model call metered and attributed, with spend over time and by module, so spend is visible before the invoice arrives.
Documents, public data, spatial data, ticketing systems, models, and the reports at the end are all within reach of a single conversation. Each capability is a tool the assistant can call, so new ones arrive without new screens.
Sources, approvals, and costs are recorded as ordinary data, not as an afterthought. The record is complete enough to replay any conversation and see exactly what was done and why.
Eunice was built first for regulated operations, where an answer without a citation is not an answer. The governance is real because it had to be.
A document search, a spatial join, a chart, a risk model, a spreadsheet export: each is a tool with a declared input and output, and the assistant chooses which to use and in what order. There is no fixed pipeline and no separate application for each kind of work.
Most platforms produce an answer. Eunice also produces the evidence: what was asked, what was called, what came back, what it cost, and where every dataset came from.
Every upload, derived table, document, and export is registered once, with its lineage, duplicate detection, and grouping into Kits. An analyst finds what already exists instead of pulling it again.
Every message, tool call, input, output, and cost is recorded without truncation. Any conversation can be replayed, filtered, or exported as a report.
Ask Eunice, then check her work.
