Hyve develops applied AI systems around domain-specific information and measurable workflows. Our work spans construction document understanding, legislative data, and coordinated software agents. Hyve Core connects these efforts through shared knowledge, source tracking, and orchestration.
Document and vision intelligence
In development
We are developing a plan reader that turns an architectural plan set into quantities by trade, with the sheet each number came from. It runs on Hyve's own server and calls no outside model service.
Where the drawings do not settle something, the reader asks the customer, pointing to the place on the page. Nothing reaches a bid until a person confirms it.
Each plan set also gets an independent take-off that records the sheet, detail, method, measurement, and formula behind every line, and flags any line that differs from the first read by more than the company's limit.
Domain knowledge
Legis records where its data came from: bill and vote records from LegiScan under CC BY 4.0, some bill text from official legislature sites, and subject terms from the Congressional Research Service vocabulary.
Every AI-written bill summary keeps the model, the prompt version, and the date it was produced, so a summary can be traced and regenerated.
Each organization's knowledge base is indexed and searched only for that organization, and content is reviewed by a person before an agent answers from it.
Agent orchestration
In HyveVibe, one model plans, another implements, tests produce the evidence, and a separate model reviews the work before a person merges it. The roles stay fixed; the models behind them can change.
Each run records its workspace, who owns it, the gates it passed, and what each tool did, so a result can be checked after the fact.
When AI tools reach Hyve's products through connectors, the organization comes from the caller's sign-in, never from the request, and only read and prepare actions are offered.
Evaluation
We evaluate extraction quality and workflow behavior with defined tasks and documented evidence. When we publish results, each will state the task, the size of the test set, the date, the method, and its limits. We have not published accuracy figures for the plan reader yet.
Compute
We are evaluating GPU-accelerated training and inference for construction document vision workloads, with emphasis on extraction quality, latency, and deployment cost.