A Forecast for Every Open Claim: The API

TL;DR — The Canotera API decouples document processing from mathematical forecasting. By separating generative AI extraction from geometric machine learning predictions, we deliver traceable settlement ranges directly into your claims management system via asynchronous webhooks.

An adjuster does not want another dashboard. They live inside a core claims administration system. If an engineering team builds a forecasting tool that requires a user to open a new tab, authenticate, and manually upload a PDF, that tool has already failed. Adoption in insurance operations is a function of proximity. The forecast must exist exactly where the work happens. This is an integration problem. It requires an API capable of absorbing massive, unstructured case files and returning a calibrated mathematical prediction without breaking the host system's workflow. The technical reality of a claim file is brutal. We are not dealing with clean data payloads. A typical casualty claim involves thousands of pages of unstructured text. Pleadings, medical records, demand letters, and internal correspondence arrive as flattened PDFs, poorly scanned TIFFs, and raw text. Ingesting this data at scale requires a pipeline built for high variance. The API must accept these raw artifacts, process them asynchronously, and map the chaos into a structured format that a mathematical model can understand.

Decoupling Extraction from Prediction

The fundamental architecture of the Canotera pipeline relies on a strict separation of concerns. Generative AI is remarkably good at reading language. It is fundamentally incapable of performing reliable mathematical forecasting. We use large language models strictly as extraction engines. When a claims system pushes a new document to our endpoints, the generative layer reads the text. It structures the unstructured. It identifies the plaintiff firm, isolates the specific medical diagnoses, maps the timeline of treatment, and extracts demands. It stops there. It does not guess the settlement value. Prediction is the exclusive domain of our geometric machine learning models. Once the generative layer compiles the extracted facts into a standardized vector, it passes that structured payload to the forecasting engine. This mathematical model operates entirely on historical precedents. It maps the geometry of the current claim against a massive, high-dimensional space of resolved cases with known outcomes. By comparing the new vector against historical resolutions, the model calculates the actual forecast.

This decoupling solves the latency and traceability problems inherent in monolithic AI applications. Large language models are slow and non-deterministic. By restricting them to an initial extraction phase, we keep the inference latency of the actual forecast under tight control. More importantly, this architecture guarantees traceability by design. The mathematical model produces a calibrated settlement range, an escalation probability, and a reserve delta. Because the input vector was built by the extraction layer, every driver behind those numbers contains a pointer back to the exact page and paragraph in the source document. The API returns the math, the drivers, and the citations in a single response.

Asynchronous Processing and Security

Processing a massive medical file is not a synchronous operation. You cannot hold an HTTP connection open while a pipeline parses gigabytes of scanned images. The Canotera API relies on an asynchronous, webhook-driven architecture. The client system submits a batch of documents and receives a job identifier. Our infrastructure takes over, scaling compute resources dynamically to handle the optical character recognition, entity extraction, and vectorization. When the forecasting model completes its inference, our system fires a webhook back to the client. This payload contains the full calibrated output ready for immediate ingestion. Security dictates every decision in this pipeline. Claims data contains highly sensitive personally identifiable information and protected health information. The API enforces strict TLS encryption in transit and AES-256 encryption at rest. We isolate tenant data logically and physically where required. The processing pipeline is designed to be as stateless as possible. Once the generative models extract the necessary structural facts, the raw documents are purged from the active processing cache. We retain only the vectorized representations necessary to maintain the forecast and support the traceability pointers.

Routing the Output into the Workflow

The value of the API depends entirely on the payload it returns. A single point estimate is useless in litigation. The Canotera API returns a calibrated settlement range. It outputs a specific reserve delta, comparing the model's projected exposure against the current reserve logged in the claims system. It provides a list of comparable resolved cases. It delivers a discrete escalation probability score. This structured JSON payload allows the host claims system to automate critical routing decisions. If the API returns a massive reserve delta and a high escalation probability on day one, the host system can automatically route that claim to a senior litigation specialist. It triggers an immediate review of defense spend allocation. The adjuster opens their familiar system and sees the forecast, the drivers, and the comparable cases already populated. They negotiate from a position backed by empirical data, not gut instinct. They detect the markers of social inflation and third-party litigation funding before the case spirals into a nuclear verdict. The API handles the complexity in the background, surfacing the exact mathematical reality the adjuster needs to make a decision.

We built this infrastructure because manual forecasting scales linearly, but litigation risk scales exponentially. You cannot hire enough adjusters to read every page of every file on the day it arrives. The API bridges that gap. It absorbs the unstructured chaos of reality and returns a structured, traceable, and calibrated assessment of risk. Math belongs in the workflow, not on a pedestal.