README.md

metadata

title: Mobility Foundation Model — LFM2.5 Co-Pilot
emoji: 🚗
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.14.0
app_file: app.py
pinned: false

Mobility Foundation Model — Co-Pilot Demo

Reference implementation of a motor insurance foundation model built on the LFM2.5-350M backbone. One backbone, three motor surfaces + text Q&A, demonstrated on the same policyholder. The architectural pattern Liquid AI is proposing to European motor insurers (Allianz Connected Platforms) as the foundation for telematics-driven underwriting and claims processing.

What this demo shows

Four tabs over one shared LFM2.5-350M backbone, all running CPU inference:

• Policyholder Trajectory — pick a policyholder from auto-discovered archetypes, see their most-recent events as a colored timeline, and one click runs all three surfaces (crash risk, claim integrity, renewal retention) in series.
• Ask the Model — type a prompt and see text generation from the same frozen backbone that processes structured events. Demonstrates the multimodal architecture: one backbone, two modalities, zero extra params.
• Why this matters — three use cases in motor insurer P&L vocabulary: loss ratio, SIU recovery, retention rate.
• How this fits your stack — VPC-native, MCP integration, license the architecture not the hosted model.

Surface performance (V7 ship candidate)

Surface	Test AUC	95% CI	XGB baseline	FM/XGB
Crash risk	0.868	[0.833, 0.901]	0.866	100%
Fraud	0.999	[0.999, 1.000]	0.871	115%
Renewal	0.734	[0.725, 0.743]	0.747	98%

Cast

Six auto-discovered archetypes from the test split:

Archetype	Description
young_risky	Age 18-25, crash events, high telematics risk
luxury_stable	Luxury vehicle, clean record, low risk
high_claims	Multiple FNOL events, frequent claim chains
fraud_flagged	Synthetic fraud-injected pattern
renewal_risk	Renewal non-accept after premium repricing
clean_commuter	Economy car, long tenure, no incidents

Architecture

event_features (1, 128, 16)  -->  EventEncoder  -->  (1, 128, 1024)
static_features (1, 10)          (attention pool         |
                                  + 6 markers)            v
                                               LFM2.5-350M (frozen)
                                               + LoRA r=16
                                                    |
                         +-------------+-------------+
                         v             v             v
                   Crash Head    Fraud Head    Renewal Head
                   (binary)      (binary)      (binary)

text prompt  -->  TextEncoder  -->  same backbone  -->  LM head (weight-tied)
                  (0 params)       (shared LoRA)

Honesty caveats

• Synthetic data. Distributions from Jeong et al. (arXiv 2102.00252). Architecture demo, not a production benchmark.
• Synthetic fraud. AUC 0.999 is on injected patterns. Production trains on the insurer's SIU labels.
• Text generation is not fine-tuned. The Q&A tab shows architectural capability (one backbone, two modalities), not domain expertise.

Read more

• LFM2 technical report (arXiv 2511.23404)
• LFM2 model weights on Hugging Face