Jung RamaDenpasar; --:--:-- GMT+8

One platform for the whole AI lifecycle.

Confide is an MLOps and AI quality management platform by Corpy&Co. — one place to take a model from data to training, evaluation, and operations, with QA and explainability.

Role
Software engineering
Type
Platform
Domain
MLOps
Year
2025

Overview

Shipping a model is easy. Operating one you can trust isn't.

Most ML work dies in the gap between a notebook and production: data scattered across folders, transforms nobody can reproduce, models with no record of what they were trained on. Confide closes that gap by making the whole lifecycle — data, training, evaluation, operations — run through one platform where everything is versioned and traceable.

And because it serves quality-critical industries, deployment isn't just on cloud, releases ship to internal staging and production, to offline client machines.

Dataset management

The lifecycle

One model, from raw data to a monitored deployment.

  1. 01

    Datasets come in, organized

    Data management

    Assets are uploaded, versioned, and tagged into annotation sets — a browsable, searchable dataset instead of loose folders.

  2. 02

    Data gets labeled — by hand or by prompt

    Annotation

    An embedded Label Studio editor for manual work, and prompt-based auto-annotation for labeling whole datasets in one run.

  3. 03

    The data gets judged first

    Analysis

    EDA scores every dataset on completeness, consistency, and accuracy, with clustering analysis and AI-generated insights before any training starts.

  4. 04

    Models train and get evaluated

    Model management

    Training runs produce confusion matrices, per-class PR curves, and best/worst class breakdowns — every model version tied to the data that produced it.

  5. 05

    Releases ship, quality stays watched

    Operations

    Releases ship as versioned container images — to staging, production, or offline client machines.

Inside the platform

The four rooms a dataset walks through.

Annotation — by hand or by prompt

Manual labeling runs in an embedded Label Studio editor, right next to the dataset. For the bulk work, prompt-based auto-annotation takes a text description of what to detect, runs a model across the set with a confidence threshold, and queues everything for human review — labeling goes from per-image work to a supervised batch job.

Prompt-based auto-annotation
Annotation editor
Asset detail

Analysis that judges the data before training does

Every dataset gets scored — overall quality, completeness, consistency, accuracy — and a clustering analysis shows how the data actually distributes. AI-generated insights read the plots for you: well-separated clusters, balanced classes, and what that means for the model you're about to train.

EDA — clustering analysis with AI insights

Defect samples on demand

Anomaly detection has a data problem: factories produce thousands of good parts and almost no defective ones. The anomaly generation tool composites real defect patches onto OK images with Poisson blending — position, size, brightness, and blending all tunable — turning a handful of real defects into a training set.

Anomaly (NG) image generation

Evaluation you can interrogate

A training run doesn't end with one accuracy number. Results come back as a confusion matrix, confidence distributions, overall and per-class precision-recall curves, and a blunt best-and-worst class ranking — with evaluation and XAI runs one click from the same screen.

Training evaluation

Three modules

Independent modules, one lifecycle.

DM

Data Management

Manage, analyze, and process data — exploratory analysis, preprocessing, and versioned pipelines that keep the lineage of every transform.

MM

Model Management

Manage, train, and evaluate models, so every version has a traceable history of what it was trained on and how it scored.

XAI

Explainability

Explainable-AI tooling built into operations, because in quality-critical deployments 'the model said so' isn't an answer.

Architecture

A control plane, an execution plane, and artifacts in between.

The web and API layer is the control plane: it registers each module and holds the metadata records. Long-running work never blocks a request: pipeline runs, training, and evaluation execute on a fleet of async workers. Everything that crosses an environment boundary is an artifact — versioned container images in a registry, pipeline configs in the database, and lineage logs tying every model version to its dataset snapshot and deployment event.

Control plane

The API and metadata layer, with modules registered per deployment.

Execution plane

Async workers for pipeline runs, training, and evaluation — nothing heavy blocks the app.

Artifacts

Container images as the release unit, versioned pipeline configs, full lineage logs.

Engineering decisions

Four things the platform has to get right.

Pipelines are versioned database objects

A preprocessing pipeline isn't a script someone ran once — it's a config stored in the database with a version, a parent it can inherit from, and the dataset it belongs to. That makes every run re-runnable and answers the lineage question: which transforms, in which version, produced this data.

One AI core, installable in pieces

The AI core is modularized into an installable package with optional add-ons per module — install only what a deployment needs. Missing features degrade to 'feature disabled' instead of a crash, and client installs stay small because they only carry the modules they use.

The release unit is an image, not source code

Merges trigger the test pipeline; a passing build produces a versioned container image pushed to a registry. Staging and production run containers only — no source code on the box. For offline client machines, the same images travel as saved archives and load without touching the internet.

Design for the smallest machine, not the biggest

Realistic datasets killed pipeline workers on 16GB machines — images loaded eagerly into memory, transforms accumulated in lists. The fix is architectural: pass paths instead of arrays, load lazily inside transforms, stream per-asset, and write to temp files. Client hardware is a constraint, not an afterthought.

Outcome

AI that quality-critical industries can actually operate.

The measure of an MLOps platform is what its users stop doing: no more loose folders of images, no more untraceable models, no more hand-run scripts nobody can reproduce.

End to end
One platform, whole lifecycle
Data intake, labeling, analysis, training, evaluation, and release run through one system — a model never loses the trail back to the data that made it.
Offline-ready
Runs where clients need it
Releases load onto air-gapped client machines and resource-constrained hardware, not just cloud servers — deployment reality drove the architecture.
Explainable
Decisions you can defend
Confusion matrices, per-class curves, and XAI runs are one click from every training result, because 'the model said so' isn't an answer in these industries.

Next project

Content

Confide Content

Confide · 2025