---
title: "Model training for anonymization and ISA | Marsstein"
description: "Our own models for anonymization and ISA validation, retrained on your footage inside the EU, shipped only after their gates and a sign-off."
url: https://www.marsstein.ai/en/infrastructure/model-training
language: en
---

# Models trained on _your own data._

We train our own models for anonymization and ISA validation. When a model meets footage it does not know yet, we retrain it on that footage inside the EU, and it ships only after it passes gates written before the run and a person signs it off.

[Talk to us](https://www.marsstein.ai/en/contact?topic=infrastructure) [Explore the anonymization model](https://www.marsstein.ai/en/infrastructure/model-training/anonymization)

AI infrastructure › Model training

## The models we train.

- Service

    ### Anonymization modelAEON

    Faces and plates in camera data from vehicles and robots, masked in the EU, with evidence per file.

    Learn more

    <https://www.marsstein.ai/en/infrastructure/model-training/anonymization>

- Service

    ### ISA validation model

    ISA validation under Delegated Regulation (EU) 2021/1958: speed limits from two independent sources, the vehicle graded automatically.

    Learn more

    <https://www.marsstein.ai/en/compliance/automotive/isa>

## How a model gets better, and how it ships.

Every release goes through the same loop. A model that fails a gate is recorded as rejected, with the reason.

- 01

    ### Harvest

    Frames and detections are sampled from production batches.

- 02

    ### Audit

    Two teacher models label every frame. Where they disagree, every frame is reviewed by eye.

- 03

    ### Dataset

    Verified false detections become hard negatives, verified misses become positives. Splits are made by recording, with provenance hashes.

- 04

    ### Retrain

    Training starts from the exact production checkpoint, proven by its hash. The gates are written before the run.

- 05

    ### Gates

    False detections on held-out data must drop, plates must hold, the full dataset is swept and other domains are checked.

- 06

    ### Sign-off

    A person signs the model off before anything ships.

- 07

    ### Release

    Weights with checksums and a model card, kept beside every earlier version.

- 08

    ### Deploy

    A model swap and engine rebuild take about two and a half minutes. Rollback is one file.

## What fixing a domain looks like.

- ### The problem

    On footage from a new region, the model in use found 1,072 faces that were not there in one recording: sand, grilles, signs and wheels.

- ### The data

    26,695 frames from that region, labelled by teacher models, every face reviewed.

- ### The first run

    Rejected at the gates: it lost 61 % of the plates.

- ### The fix

    Plates relabelled by the model in use, false detections added as hard negatives four times over.

**−94 %** false detections on the new footage.

**+0.4 %** plates found.

**None** drift on European footage. The model was approved.

## Where training runs.

- ### Apart from production

    Training has its own resources in an EU region. Production only reads released models; nothing is pushed into it directly.

- ### In the EU

    Customer footage never touches GitHub. All data stays in the EU.

- ### Every version on record

    Which model runs, which ran before, and every rejected run with its reason.

- ### Pinned by checksum

    The products that use a model pin it by its checksum.

## Also from Marsstein

- ### Anonymization modelAEON

    Faces and plates in camera data from vehicles and robots, masked in the EU, with evidence per file.

    Learn more

    <https://www.marsstein.ai/en/infrastructure/model-training/anonymization>

- ### Data center

    Runs in the Marsstein EU cloud: your footage lands in an EU bucket and is masked inside the EU.

    Learn more

    <https://www.marsstein.ai/en/infrastructure/data-center>

- ### On-premise

    Marsstein on your own servers: in your own cloud account or on your own GPU servers.

    Learn more

    <https://www.marsstein.ai/en/infrastructure/on-premise>

## Send us ten minutes of footage.

[Talk to us](https://www.marsstein.ai/en/contact?topic=infrastructure)
