Insurance runs on documents. Riska reads and writes them.
Everything that arrives is classified, split and extracted onto your data model. Everything you owe is generated back out of it. Every value keeps the page and the box it was read from.
Every file takes the same path. Classification, splitting and extraction run as one automated pass, with a named stage for each decision the models make. Open any stage, see what it saw, restart from there.
Classify.
A light first pass names the file against the document types you defined, before a single field is read.
Route to the product.
The submission is matched to the program it belongs to, scoped to the products you allow it to choose from.
Split, then extract.
A packet is carved into named, page-ranged sections, and each section is read against the fields mapped to it.
Park when unsure.
A stage that cannot resolve something opens a task, assigns it to the right person and waits. It never guesses.
“What changed my mind was clicking a number and landing on the exact box it was read from.”
Nothing is taken on trust. Extraction produces evidence, not facts. Every value carries where it was read and how sure the model was, and what becomes a record is still a decision someone makes.
Click the value, land on the box.
Every extracted value keeps the page it came off and the coordinates of the box it was read from. The viewer draws that box back over the original, so a disputed figure is one click from its source rather than a search through an attachment.
A score beside every field.
Each field comes back with its own confidence, colour-coded in the fields panel next to the document. The weak ones are the ones a reviewer sees first, so attention goes where the model was least sure instead of down the page in order.
Disagreements survive.
When two files answer the same question differently, both candidates are kept. You see which value won, that it won on confidence, and what lost. Nothing is quietly overwritten by whatever arrived last.
Graded before it ships.
Attach the values you expect to a sample and run the whole pipeline against it. Every stage gets a pass or a fail, the run never touches your data, and the prompts you tune are versioned and released like the rest of your configuration.
Add a document. Add a field. A document Riska has never seen is configuration, not a project. The types, the criteria that tell them apart and the fields they fill are yours to define, and everything they capture joins the same Universal Context the rest of the platform reads from.
Any document, not a fixed list
Name the documents your programs actually receive and write the criteria that tell them apart. Scope one to a single product or to everything you write.
The mapping proposes itself
Point the setup at a few samples and it returns a field-by-field mapping onto your objects, prompt included. Nothing is written until you apply it.
New fields become new risk factors
An extracted field lands on your own model, so appetite rules, rating inputs, dataviews and the assistant can all use it the moment it arrives.
Riska writes them too. Upload a template, tag it against your data model, and the documents you owe go out filled from the record they describe. Binders, policy forms, quotes and notices come off the same graph that captured the submission.
Tag it once.
Upload your own template and mark the places your data belongs, editing it in the browser. An agent reads the document first and proposes where each token and each repeating section should go, and why. You accept the suggestions you want.
Any attribute on your model.
A token can pull any attribute you have, followed across relationships, and a loop expands a list row by row. The mapping is resolved and validated against the template before a run ever reaches a customer.
Generated where the work is.
Pick the record, and the templates that apply to it are the ones you are offered. The finished document is filed against that record and audited like everything else, or produced by a workflow at the moment the policy binds.
Documents are where the work starts.
What the pipeline reads is what triage sorts, appetite scores and the assistant answers from. Read once, used everywhere.
Bring the messiest file you have.
Send something out of your own book that nobody has managed to automate. We will classify, split and extract it live, open every citation you ask about, and generate a document back out of it.
About 30 minutes, and you keep the extraction.