e Euler ecosystem interactive field guide

From raw files to defensible findings.

Euler gives file-based ML projects a shared language as data moves through indexing, loading, transformation, training, inference and evaluation—while each tool stays useful on its own.

Follow an experiment See the closed loop

6composable tools
in the core path

1shared on-disk
dataset contract

0required remote
tracking services

euler://fog-robustness portable

research question / 01 Does controlled fog augmentation improve depth robustness?

sourceRGB · depth · segmentation · K

files + zip

{}

identitydataset-head.json + stable IDs

indexed

↗

provenanceconfig · Git · environment · metrics

local-first

✓

claimpaired metrics + per-file evidence

reviewable

LDalign by ID

EVhonest metrics

VWorchestrate

01 / why

The hard part of ML experimentation is not calling model(x). It is preserving the identity, meaning and provenance of x across every hand-off.

The churn test

Change one thing.
See who has to move it.

Pick a change. Watch it ripple.

× Manual glue 6 touchpoints edit · run · discover mismatch · repeat

edit

Loadsplit.csv

edit

Derivesubset flag

edit

Traintrain.yaml

edit

Inferinfer.json

edit

Evaluategt paths

edit

Inspectview filter

duplicate state order-dependent stale result risk

✓ Euler contract 1 touchpoint change once · every consumer resolves it

{}

dataset contractsplit = fog_test_v2

1 edit

01Loadsynced

02Derivesynced

03Trainsynced

04Infersynced

05Evaluatesynced

06Inspectsynced

identity travels semantics travel outputs re-enter the loop

The evidence chain

One question. Seven stages. Two workflows.

Interactive experiment

Follow one frame from raw file to research claim.

Demo question: Does physics-based fog augmentation improve monocular depth robustness?

1 reusable identity

scene_01 / frame_00042 stage 00

/data/vkitti2

├─ rgb/Scene01/00042.png

├─ depth/Scene01/00042.png

├─ classSegmentation/…

└─ textgt/Scene01/intrinsics.txt

Research question Will the model hold up as visibility falls?

Start with the files you already have

your project

Frame a falsifiable research question

01 / input

Folders or ZIP archives from a simulator, sensor rig or benchmark

02 / work

Choose the modalities, conditions and held-out scenes that make the claim testable

03 / evidence

A concrete question with explicit inputs and a comparison target

/data/vkitti2/{rgb, depth, classSegmentation, textgt}

Euler begins at the filesystem boundary; it does not require data migration into a proprietary store.

The closed loop

One path in.
EL datasets everywhere.

// folder or .zip ↘ hierarchy-aware IDs + plug in any consumer

Raw sourcerandom files

rgb · depth · K · …

euler-loadingsame contract

RGBdepthsegK ↘

/scene01/cam0/00042

✓ loaders resolve 0 custom dataloaders

Any consumerimport euler_loading

compatible

The package boundaries

Same stages.
Different topology.

copy assumptions vs resolve contract

System architecture

One contract. Several focused tools.

Select a node to see what crosses each package boundary.

artifact flow re-indexed output control / observation

portable data plane

orchestration layer

{}

shared language

Dataset contract

IDs, modality semantics and loader instructions travel together.

dataset-head.json, index.json and split artifacts are the stable hand-off between independently useful tools.

reads

Dataset identity, hierarchy, properties, units and addon metadata

writes

A portable description every downstream package can understand

↳Semantics travel with data

Adopt progressively

Use the smallest ecosystem that solves today’s problem.

Add one layer when it removes one recurring chore.

I only need reliable multimodal loading

ds-crawlereuler-loading

Index once. Keep your existing model and loop.

I am building a controlled data study

+ euler-preprocess

Generate variants without losing source identity.

I need reproducible model evidence

+ euler-traineuler-inferenceeuler-eval

Record, predict and score in one declared space.

A team needs to run this on a cluster

+ Euler View

Dispatch, observe and compare the same portable artifacts.

What this buys your methods section

“We paired modalities and predictions by stable dataset IDs, preserved dataset-native semantics through preprocessing and inference, recorded run provenance locally, and evaluated all models against the same indexed ground truth.”

From raw files to defensible findings.

Change one thing.See who has to move it.

One question. Seven stages. Two workflows.

Follow one frame from raw file to research claim.

Frame a falsifiable research question

One path in.EL datasets everywhere.

Same stages.Different topology.

One contract. Several focused tools.

Dataset contract

Use the smallest ecosystem that solves today’s problem.

I only need reliable multimodal loading

I am building a controlled data study

I need reproducible model evidence

A team needs to run this on a cluster

Change one thing.
See who has to move it.

One path in.
EL datasets everywhere.

Same stages.
Different topology.