Three hurdles teams face when building vision AI in‑house

01

Data cannot be taken outside

Production line images, facility videos, and inspection results are the company’s core assets. To upload them to a cloud training service, you first have to pass security reviews, and in many cases you simply can’t.

02

Tools are scattered

Labeling is done in a separate program, training is run by personal scripts, experiment logs live in Excel, and model files sit in a shared folder. When the owner changes, it becomes impossible to track which data was used to build which model.

03

No one to own the infra

To wire together and operate GPU drivers, training environments, experiment tracking servers, and storage, you need a dedicated engineer. In small teams, the AI engineer ends up taking on that work as well.

miniOps’ answer

Install on your existing GPU server and keep data on‑premise, while handling everything from labeling to model management in a single interface.

Five stages connected in one interface

Seamless from data registration to management of models deployed in the field

STEP 1

Dataset Hub

Register images and labels with a single ZIP upload, and check class distribution and review status.

STEP 2

Labeling & Review

Draw boxes and polygons in the web editor, let AI auto‑annotate drafts, then have humans review them.

STEP 3

Training

Start training directly from the dataset screen, and monitor progress, loss curves, and resource usage in real time.

STEP 4

Result Analysis

Review training curves, PR curves, confusion matrices, and metrics, and download weights.

STEP 5

Model Registry

Register models by version and manage them across Staging and Production stages.

✓

Installed, single‑server setup

Install on a single server using Docker Compose. All images, labels, and model files are stored on a designated internal disk.
✓

All you need is a web browser

Access over the internal network with no extra programs to install. The interface is provided in Korean.
✓

Built on proven open source

We integrate widely used open‑source components like MLflow, MinIO, and PostgreSQL into a single interface. You’re not locked into any specific vendor.
✓

Object detection & segmentation

Build image‑based inspection models such as defect detection, parts recognition, and region segmentation.

01 · Dataset Hub

Manage scattered images in one place

✓

Register via ZIP upload

Upload a ZIP file containing images and labels, and the format is detected automatically. Supports COCO JSON and YOLO TXT, and you can also upload images without labels.
✓

Automatic thumbnail generation

Preview images are generated at upload time, letting you quickly skim through even thousands of data items.
✓

Status at a glance

Each card shows processing, ready, and failed statuses, as well as whether labels exist. Filter by task type and status.

miniOps dataset hub screen — Dataset Hub: shows status, image count, and classes per dataset as cards

Quantify the state of your data

✓

Check class imbalance

Compare label counts by class in bar charts to identify under‑represented classes before training.
✓

Track review rate

See the ratio of images whose review has been completed. You can also train using only the reviewed subset.
✓

Class filter

Filter and inspect only the images that contain a specific defect type.

miniOps dataset statistics screen — Stats tab: number of annotations, counts by type, review rate, class distribution

02 · Labeling

Label directly in your browser

✓

Boxes and polygons

Draw bounding boxes for object detection and polygons for segmentation.
✓

Class management

Add and select classes inside the editor. Classes that are in use are protected from accidental deletion.
✓

Review flags

Work through images in sequence and toggle each image between “Pending review” and “Review complete” via button or R keyboard shortcut.

miniOps annotation editor screen — Annotation editor: tool selection, class management, and label list in one screen

AI drafts, humans review

✓

Text‑driven automatic annotation

Enter the names of targets as classes, and an AI model finds those objects in images and creates boxes and polygons. Usable out of the box with no prior training.
✓

Batch processing per dataset

Run auto‑annotation in bulk only for unlabeled images, only for unreviewed images, or for all images. Control strictness using a confidence threshold.
✓

Notifications when done

When the job finishes, the counts of successes and failures appear in the notification center. You can keep working on other tasks in the meantime.

Note: Automatic annotation accuracy varies depending on the target and image characteristics. It works well for general objects; highly specialized targets such as tiny defects still require human review and correction.

miniOps automatic annotation result — Result after specifying truck, forklift, and car as classes. Draft annotations created by AI.

03 · Training

Start training right from the dataset screen

✓

Minimal input required

Just set the model name, task type (object detection / segmentation), and number of epochs. Batch size and learning rate can be set automatically.
✓

Automated splits and format conversion

The platform handles train/validation/test splits and conversion to training‑ready formats. No need to write scripts.
✓

Transfer learning and automatic experiment logging

Start from pre‑trained models to get results quickly even with limited data, and have hyperparameters, metrics, and model files automatically logged to MLflow.

miniOps model training start screen — Start training: opened via the “Start training” button on the dataset detail screen

See training progress in real time

Progress, loss values, precision, recall, and mAP are updated on screen at the end of each epoch. Queue, running, completed, and failed job counts are visible at a glance.

Monitor GPUs without logging into the server

Check GPU utilization, memory, temperature, power usage and last‑hour trends per GPU, along with running processes and driver / CUDA versions, all from the web UI.

04 · Results

Evaluate model performance with evidence

✓

Automatic visualization

Once training completes, instantly see training curves, F1 / PR curves, and confusion matrix images, and download them for reports.
✓

Download weights

Download both the best‑performing model (best) and the final model (last).

miniOps confusion matrix screen — Confusion matrix: shows per‑class hits and confusions

05 · Model Registry

Keep a record of which models are in production

✓

Versioning

Each registration under the same name increments the version, and you can compare metrics by version in a table.
✓

Stages

Classify model status into four stages: None, Staging, Production, Archived. When a model is promoted to Production, the previous Production version is archived automatically.
✓

Download history

Keep a record of who downloaded the model file and when.

miniOps model detail screen — Model detail: version selection, stage management, training metrics, and hyperparameters

Your data never leaves your servers

We bundle proven open‑source components into a single Docker Compose stack and install it on your GPU server

Customer GPU server (Docker Compose)

Nginx reverse proxy

Single access endpoint. All requests pass through here.

Web UI

Next.js

API server & training engine

FastAPI

PostgreSQL

Datasets, labels, job logs

MinIO

Image & model file storage

MLflow

Experiment tracking & model registry

On‑premise disk (designated path)

All data is stored under a single folder on the host disk, making backup and migration straightforward.

On‑premise installation

No cloud account required. Images and models are stored only on the disk of the server where miniOps is installed.

Storage stays inside your network

The image store (MinIO) is never exposed directly to the outside; images are only served via the server.

Single entry point

Provide the web UI, APIs, and experiment tracking UI under a single address.

Management scripts included

Start, stop, restart, and check status with a single management script command.

Key specifications

Category	Item	Details
Functionality	Training tasks	Object detection, instance segmentation
Functionality	Data formats	COCO JSON, YOLO TXT, unlabeled image ZIP (JPG, PNG)
Functionality	Annotation	Bounding boxes, polygons, AI automatic annotation
Functionality	Experiment & model management	MLflow‑based logging, versions & stages (None · Staging · Production · Archived)
Functionality	Notifications	Training start / completion / failure, automatic annotation results
Environment	Installation method	Docker Compose (single server), installation script provided for Ubuntu 24.04 LTS
Environment	Hardware	NVIDIA GPU (NVIDIA Container Toolkit), 16 GB+ RAM recommended
Environment	Software	Docker 20.10 or higher, Docker Compose 2.0 or higher
Environment	Access & language	Web browser (Chrome recommended), Korean

Conditional automatic retraining, inference APIs, and deployment history & approval management are currently under development; timelines will be provided separately.

Implementation process

We support you from consultation through to operations

01

Consulting & demo

We review your inspection targets and data status, and walk through the workflow using the actual screens.

02

Environment check

We check your existing GPU server specs and network topology, then plan the installation.

03

Installation & data registration

We install on your server and register your existing images and labels as the first dataset.

04

Training & operations support

We provide hands‑on training from labeling to model registration and respond to questions during operations.

Get a demo using your own data

Share your inspection targets and sample images, and we’ll prepare a demo tailored to your data.

Request a demo contact@techniflows.com

On‑premise

No external data transfer

5 steps

Connected in one interface

1 server

Installed with Docker Compose

Open source

No vendor lock‑in

The screenshots were taken using the public dataset DeepPCB (MIT License) and public domain / CC0 photos from Wikimedia Commons. Features and screens are subject to change by version.