Nouli Insurance Loss Event: Property-claim intake

How it works

An initial claim narrative goes to the loss-event classifier, which returns a category and probabilities. A deterministic Nouli routing rule then selects an intake queue or review.

The model identifies the loss event from the narrative. A separate, hand-authored rule maps its answer to a queue. The included Nouli workflow connects the two steps and records their execution.

Model at a glance

Property	Value
Input	Initial property-claim narrative in English
Output	One of 22 loss events, probabilities and an illustrative queue
Categories	Water damage, fire, theft, weather events and other property-loss causes
Runtime	Local CPU or CUDA; CPU rechecked for this release
Base model	RoBERTa-large
License	Apache-2.0; upstream MIT notice applies

Example input and output

Input: “Sewage backed up through the bathroom drain and damaged the floor.”

Recorded output (probabilities and trace fields omitted):

Recorded output

{
  "answers": {
    "loss_event": {"choice": "Backflow waste"}
  },
  "routing": {
    "queue": "water_damage"
  }
}

The specialist identifies sewage backflow, then the routing rule assigns the water-damage queue. This example was run with the supplied weights on CPU. Input · Complete output.

Examples for claims handlers and loss adjusters

These constructed first-notice-of-loss (FNOL) examples cover all 22 model categories. Each result was recorded on CPU. Complete inputs, probabilities and queue outputs.

For an intake handoff, retain the original notice alongside the suggested event and queue. The reported cause matters: a broken window can follow hail, forced entry or an accidental impact. Policy coverage and settlement require a separate assessment.

Water-related notices5 examples

Example queue: water_damage.

Reported FNOL	Recorded loss event
During prolonged heavy rain, water entered through a poorly sealed roof joint and stained the bedroom ceiling.	Atmospheric precipitation
Sewage backed up through the bathroom drain and damaged the floor.	Backflow waste
The nearby river overflowed its banks and floodwater entered the ground floor of our home, damaging flooring and furniture.	Flood
Groundwater rose beneath the house after weeks of wet weather and seeped through the basement walls and floor.	Groundwater
A pipe beneath the kitchen sink burst, releasing water across the floor and damaging the lower cabinets.	Waterleakage

Theft, forced entry and malicious damage4 examples

Example queue: crime_damage.

Reported FNOL	Recorded loss event
Our home was broken into while we were on holiday. The intruders smashed a rear window to enter and took jewellery from an upstairs bedroom.	Burglary
A person threatened me with a knife outside my front door and forced me to hand over my handbag and phone.	Robbery
My bicycle was taken from the front garden while I was inside. No doors, locks or windows were damaged.	Theft
Someone deliberately sprayed paint across our outside wall and smashed the garden lights. Nothing was taken.	Vandalism

Weather and external damage8 examples

Example queue: weather_or_external_damage.

Reported FNOL	Recorded loss event
Lightning struck our house chimney directly and damaged the chimney masonry and roof tiles.	Direct lightning strike
An earthquake shook our home and left fresh cracks across the external walls and foundation.	Earthquake
A large tree fell onto the garage during the night, crushing the roof and damaging the guttering.	Falling trees, poles and other objects
Large hailstones struck the house during the afternoon and cracked several roof tiles and the conservatory glazing.	Hail
The slope behind our house collapsed in a landslide, pushing soil and rocks into the garden wall and outbuilding.	Landslide, rockfall
Strong winds lifted roof tiles and tore the guttering from our house during a windstorm.	Storm
A heavy accumulation of snow and ice overloaded the conservatory roof, causing it to sag and collapse.	Weight of snow or ice
A wild boar broke through the garden fence and damaged the planted beds and exterior wall of the shed.	Wild Animals

Fire and explosion2 examples

Example queue: fire_or_explosion.

Reported FNOL	Recorded loss event
A fire started in the kitchen and spread to the cabinets, leaving burned surfaces and smoke damage throughout the room.	Fire
A gas cylinder exploded in the detached garden shed. The blast damaged the shed walls and nearby windows.	Explosion

Glass, electrical and appliance failures3 examples

Example queue: glass_electrical_or_breakdown.

Reported FNOL	Recorded loss event
While moving a cupboard, I accidentally struck the living-room window with its corner and shattered the pane.	Glass Damage
An electrical surge from the supply network damaged the television and refrigerator. An electrician confirmed the damage followed a voltage spike.	Overvoltage
Our washing machine stopped working because its internal motor failed during normal use. The engineer found a mechanical breakdown inside the appliance.	Extended warranty

The same damaged item can belong to different event categories. These examples illustrate the fixed taxonomy; performance on real insurer notices remains unmeasured.

Quickstart

Use Python 3.12 and the HF CLI. Download the complete repository:

Terminal

hf download nouli-ai/nouli-insurance-loss-event --revision v1.0.1 --local-dir ./nouli-insurance-loss-event
cd nouli-insurance-loss-event
python3.12 -m venv .venv
.venv/bin/python -m pip install -r requirements.txt
.venv/bin/python predict.py --checkpoint . --device cpu \
  --text-file examples/notice.txt

Replace the example file with your own initial notice. The helper applies the release preprocessing and returns all 22 probabilities. No hosted API key or separate base-model download is required.

To run the two-step workflow through Nouli CLI 0.5.1:

Terminal

mkdir -p outputs
.venv/bin/python nouli_demo.py --checkpoint . \
  --nouli "$(command -v nouli)" --device cpu \
  --text "$(cat examples/notice.txt)" --output outputs/workflow.json

See setup and integration for CUDA, offline use, the routing rule and output fields.

Base model and data

Based on FacebookAI/roberta-large, with data from Gratex’s GNOTHEIA Synthetic Insurance Dataset v2. The dataset contains synthetic household and building claim scenarios. Source revisions and provenance.

The repository includes the complete selected classifier, tokenizer and local runtime. The workflow and queue mapping are application rules supplied with the example.

Limitations

Validated on synthetic English data. Performance on real insurer notices remains unmeasured.
Supports a fixed 22-category taxonomy. It always chooses a category for a valid narrative; unfamiliar or closely related events can be misclassified, including theft and burglary.
Inputs beyond 384 tokens are truncated. Keep the reported cause and circumstances near the beginning.
The 0.8 review threshold is illustrative; model probabilities do not guarantee correctness. Validate thresholds for your workflow.
Classifies the reported event; coverage, fraud, pricing and payout decisions are outside its scope.

License and citation

Apache-2.0, with the upstream MIT license retained in BASE_MODEL_LICENSE.txt. See NOTICE for attribution.

Nouli AI. Nouli Insurance Loss Event. Cite the published version and commit.