We built Gemina, AI document APIs that extract and classify files, then let products search them and chat with them, with citations to the source.
Client: GeminaIndustry: SaaS & TechnologyService: AI & Automation
Overview
Client
Gemina
Document AI APIs for product teams and SaaS platforms
Engagement snapshot
The challenge
Plain OCR only returns text. A product team still has to work out which number is the invoice total, store the result, make it searchable and answer questions about it.
Calling a language model directly helps with single answers, but it leaves storage, search and security to build. Asking a model to add up figures also carries a risk, because a total can look right and still be wrong.
Raw OCR output cannot feed an accounting system or a workflow until someone maps it to fields.
Totals generated by a model, rather than calculated, are hard to trust in finance work.
A chat reply that cannot point back to a document leaves users checking it by hand.
The solution
We built extraction that returns document fields as typed JSON, including headers, line items and custom schemas, with a confidence score on each field.
Documents can be found by meaning or by field, and sums, averages and counts run as database queries over the extracted data.
Users ask questions in plain language, and each answer links back to the documents it came from.
Key features
Upload a sample and AI suggests the fields to extract. Teams can then add, remove or rename them and call the template by ID.
Business impact
Your own schema
FileTag labels PDFs and images with metadata such as vendor, date, type and total, suggests filenames and offers an enriched copy to download.
Business impact
1,500 free tags a month
Teams can embed the source document beside editable fields, and reviewers focus on the uncertain values.
Business impact
Audited corrections
Before and after
Before
OCR text to clean up
The usual route gives a block of text that still has to be mapped to fields by custom code.
After
Fields as typed JSON
Vendor, dates, totals and line items come back named and typed, each with its own confidence.
Before
Spreadsheets or model guesses
Spend figures are normally added up by hand or requested from a model that estimates them.
After
Database queries
Totals are calculated by the database, and the response says when missing data would undercount a figure.
Before
Opening each file
Finding a clause or an amount usually means searching for the file and then reading it.
After
Cited answers
Users ask directly and follow a citation to the source document.
Results
Typed JSON output
A product can send in a document and receive fields it can store, filter and pass to other systems.
Before: Plain textAfter: Typed fields
Database-computed totals
Sums run in the database, and a request across currencies returns separate groups instead of one mixed figure.
Before: EstimatedAfter: Computed
REST, SDKs, MCP, chat UI
Developers can use REST or an SDK, connect an agent by MCP, or place the chat component inside their own product.
Before: Custom pipelineAfter: One API
Trust and control
Uncertain fields are flagged for review, so people can check them before the data is used.
Flagged for reviewCustomers set retention periods and can delete documents through the API or let purge dates remove them.
Retention settingsThe embedded chat uses scoped session tokens, so the customer's API key does not reach the browser.
Scoped tokensWhen a number would undercount because data is missing, the response says so.
Stated gapsFaster delivery
All four modules work from the same uploaded document, through the same API.
Reads documents and returns headers, line items and template fields.
Indexes documents for search and runs aggregations over extracted fields.
Produces grounded answers that cite the source documents.
Exposes extraction, tagging, search and aggregation as MCP tools on one hosted server.
Gemina turns an uploaded document into typed JSON, including headers, line items and custom fields, with a confidence score on each field. The same document can then be searched, included in calculated totals and queried through chat. Each chat answer links back to the documents it came from, so users can check it.
Gemina calculates totals with database queries over the extracted fields, not with an AI model's estimate. Sums, averages and counts all work this way. If a request spans several currencies, the server groups them separately and says so. The response also warns when missing data would make a figure undercount.
Developers can add Gemina through a REST API, one of five SDKs, a hosted MCP server with 16 tools for AI agents, or a drop-in React chat component. The component uses scoped session tokens, so end users can ask questions about documents while the API key stays out of the browser.
If your users still type figures out of invoices, contracts or forms, we can help you build AI that reads those files and shows its sources.