AI applications people can use and inspect.
Minutehouse develops AI features for company documents and incoming text. The software around the model determines how people submit material, inspect a result and correct it.
A project may focus on a document assistant, a classifier or a drafting feature. It can also combine those functions in one application when the use calls for it.
The project list shows active items. Select Archived in the status filter to include archived projects.
Open the project list and select Archived in the status filter.
Document assistants
An assistant can search a collection of documents and prepare an answer using the relevant passages. A useful screen lets the reader open those passages in the original material.
Development includes preparing documents for search, retrieving passages and deciding how the application presents them. The collection may contain long manuals, short guidance pages or several versions of the same document. Those differences affect what the search returns.
A document title alone gives a reviewer little to inspect. We can include the passage used for a particular statement, its source link and the information needed to distinguish relevant versions. When the collection does not contain an answer, the screen can identify the missing information and leave the question unresolved.
The project also defines how new, revised and removed documents enter the collection. Test questions help show what changed after an update.
Text classification
A classifier assigns a label to a message or other text. Possible applications include sorting requests for a review queue or identifying the subject of an incoming document. The categories and their definitions are part of the application design.
Begin with real examples that can be used appropriately for development. A category should explain what belongs in it and how to handle overlap with another category. Mixed requests and unfamiliar material need a defined treatment too.
Minutehouse can develop the classification feature and a review screen that keeps the original text available. Corrections made during review can identify a vague category, a missing example or a model error. Each calls for a different change.
Evaluation compares the suggested labels with the agreed examples. A headline percentage on its own can hide the category that receives most of the mistakes. Reviewing errors by category helps show where the feature needs attention.
Drafting tools
A drafting feature can prepare a reply, a summary or another piece of text from supplied material. The intended reader, required facts and output format should be clear before the prompt is written.
The application can put the input beside the draft and give the reviewer an ordinary editing surface. Where a draft uses retrieved documents, source links can remain available during editing. Sending or publishing the result is a separate application function and must be deliberately included in the assignment.
Tests should cover missing facts, contradictory input and requests outside the feature's intended use. They should also check whether the draft preserves required details and follows the requested format.
Testing AI features
Keep a set of examples with the result expected from each. For document search, inspect which passages were retrieved. For classification, compare the label. For drafting, identify the required facts and the changes a reviewer would make.
Use the same suitable examples when comparing changes to prompts, retrieval settings or a model. Record the type of error, not simply whether someone liked the output. The findings can point to document preparation, category definitions, interface behavior or the model configuration.
The application also needs ordinary software testing. Its screens, data handling and connected interfaces can fail independently of an AI result.
Project planning covers data handling, model services and hosting, along with the application features. Maintenance and support arrangements are agreed for the deployment.