See the model
Layers, data flow, skip paths, inputs and outputs are visible as a graph instead of being buried inside a codebase.
VISUAL AI LEARNING · MODEL BUILDING · REAL RUNTIME
MLB Studio
v1.0.0 BetaAI should become a literacy—like mathematics and language—not knowledge reserved for engineers.
MLBricks MLB Studio turns the structure of AI into something learners cansee, change, run and share. Build models from understandable bricks, prepare real datasets, train and validate on available compute, generate outputs, serve a model through an API, and move work between local and cloud repositories.
[ GLIMPSE OF MLB STUDIO ]
One visual workspace for bringing data in, preparing it, inspecting every step and handing it directly to the model workflow.
MLBricks Studio keeps the data path visible as a connected graph. Learners can choose a source, configure each processing brick, follow connections across the canvas and inspect the resulting dataset without leaving the workspace.
The current Data Processing view brings together the Data Library, processing graph, live Inspector and Data Repository. A Hugging Face dataset can move through text processing, train/validation/test splitting and tokenization into a prepared dataset that is ready for Model Builder.
DATA → MODELFETCH · PROCESS · SPLIT · TOKENIZE · INSPECT · PREPARE
MLBRICKS STUDIO WORKSPACEINSPECT A BRICK · PREPARE THE DATA
[ WHY MLB STUDIO ]
Move AI education from “use the tool” to “understand the system.”
Learning becomes more meaningful when a learner can connect a concept to a visible system and then experiment with it. MLB Studio is designed as a bridge from first contact with AI to real model-building practice.
Change one idea
Swap a brick, change a width, move a connection or modify runtime settings and observe what the change means.
Run real computation
Supported graphs compile into real PyTorch/MLBricks modules; training, validation and generation are connected to the Python kernel.
Grow into the details
The Inspector exposes the underlying component API. Visual building is an entry point, not a replacement for deeper technical understanding.
WHY VISUAL FIRST?
Research on visual programming has repeatedly found that reducing syntax burden can make computational ideas more accessible to novices and let learners focus more on structure and semantics. MLB Studio applies that learning principle to model architecture while keeping real APIs and runtime behavior visible.
[ AI ACROSS THE CURRICULUM ]
AI is not one subject.
It is becoming a way to work inside every subject.
UNESCO has argued that AI literacy should be developed broadly and considered across sciences, humanities and the arts. MLB Studio’s long-term educational direction is to make the same model-building language approachable from any discipline.
Make abstractions executable.
Connect functions, probability, optimization, vectors and loss to a model that can actually be changed and trained.
Work with biological data.
Use datasets as the entry point for learning how models detect structure in text, sequences, measurements or images.
Connect models to systems.
Use AI as another computational tool alongside simulation, measurements, sensors and experimental data.
Study language and information.
Explore how text becomes tokens, representations and predictions—and learn to question model outputs, bias and evidence.
Turn data into testable models.
Learn the pipeline from data preparation to model choice, evaluation, deployment and responsible interpretation.
Create with computational media.
Image and audio data workflows create a path toward understanding how generative and perceptual systems are assembled.
These are learning/application directions, not a claim that v0.7.35 already ships a finished training template for every discipline. The current real model executor is strongest for supported text language-model graphs; MLB Studio already exposes image/audio data processing and broader model components for continued expansion.
[ ONE WORKFLOW ]
From a question to a model people can actually use.
MLB Studio connects the educational path and the engineering path instead of treating them as separate products.
01Choose a questionsubject · problem · idea
→02Prepare datasource · clean · split · tokenize
→03Design the modelbricks · ports · architecture
→04Build & checkvalidate graph · data compatibility
→05Train & validateloss · samples · checkpoints
→06Generate & testprompt · sampling · outputs
→07Serve & shareAPI · web app · cloud
[ WHAT IT CAN DO TODAY ]
A learning interface on top.
A real Python runtime underneath.
The features below are grounded in the uploaded MLB Studio v0.7.35 source. Where something is a roadmap direction rather than a shipping capability, it is marked explicitly.
Build with real MLBricks components.
Drag model bricks into a left-to-right graph, use Auto Connect or manual ports, create skip/extra paths, open nested reusable layers, and inspect the installed MLBricks API from the right panel.
VISUAL ARCHITECTUREDesign the data pipeline too.
Use Hugging Face, Kaggle, URL or local sources; clean and split text, tokenize it, process image/audio datasets, create DataLoaders and register Prepared Datasets.
DATA WORKSPACERun the experiment, not a mockup.
Supported text LM graphs compile to a realtorch.nn.Module, train with optimizer/validation/checkpoints, then generate through the selected runtime configuration.
Use the compute you actually have.
Select CPU or visible GPUs, Auto/Native/PyTorch backend policy, eager or compiled execution, compile mode, and FP32/FP16/BF16 precision.
LOCAL / NOTEBOOK / SERVERKeep models, components and data reusable.
Save model layouts, custom bricks and data pipelines; keep built/imported models in Model Repository and prepared datasets in Data Repository.
REUSEPush and pull the work.
Move datasets, models or complete MLB Studio projects through Hugging Face, GitHub, AWS S3, Google Cloud Storage and Azure Blob Storage.
PORTABLE .MLBRICKS BUNDLESBring work from the machine or notebook.
Scan local/Kaggle/Colab-style paths recursively for compatible model checkpoints, MLB Studio bundles and datasets, then restore them into the repositories.
FILESYSTEMConnect the model to a device or app.
Start a local HTTP model server with browser playground, health/model endpoints, generation and OpenAI-style completion routes. Use localhost, LAN, or an optional public HTTPS tunnel.
API + WEB PLAYGROUND[ MODEL → DEVICE ]
The learning project does not have to stop inside the notebook.
A trained or imported model can become a service that another application, computer, phone or embedded system calls.
One resident runtime
The server loads the selected model and keeps it resident while requests arrive.
Same machine
Use the generated local URL from tools or apps running on the host.
Phone / local device
Call the model from another device on the same network when the environment permits LAN access.
Remote web app
Optionally expose the notebook/server through an ngrok tunnel for remote access.
HTTP SURFACE
GET /healthGET /v1/modelPOST /v1/generatePOST /v1/completionsBearer API-key protection is enabled by default. API keys and tunnel tokens are treated as session values rather than serialized project content.
[ DATA & CLOUD ]
Build here. Continue somewhere else.
AI learning becomes more practical when models and datasets are portable instead of trapped in one notebook session.
Hub-native model and dataset workflows
Push/load MLB Studio datasets, trained model packages and projects. Private repositories use locally authenticated credentials.
Portable project artifacts
Store and restore MLBricks bundles through a familiar development repository.
Object storage
Push and pull portable.mlbricks.zipbundles through S3, Google Cloud Storage and Azure Blob Storage.
Bring live data closer to the graph.
v0.7.35 does not ship a direct SQL/NoSQL database source node. A natural extension is a database/API adapter layer feeding the existing Data Processing workspace while preserving credentials outside project files.
SESSION-ONLY SECRETS
Hugging Face tokens, GitHub tokens, AWS keys, Google service-account JSON, Azure connection strings, API keys and tunnel tokens are intentionally excluded from normal MLB Studio design serialization/cloud bundles in the current implementation.
[ FOR BEGINNERS → FOR BUILDERS ]
Start visually. Reveal complexity when it becomes useful.
The goal is not to keep learners permanently inside a simplified interface. The goal is to give them a path into the real concepts without making syntax the first barrier.
Open a sample.
Use Gallery examples, inspect components, follow connections and learn what an input, embedding, mixer, FFN, residual and head actually do.
Change the system.
Edit model settings, connect a prepared dataset, choose runtime/device settings, train, validate and compare outputs.
Build your own architecture.
Create nested custom bricks, manual signal paths, reusable models and data pipelines, then export or publish the result.
Put it into another context.
Serve it through an API, connect a web app/device, or move the model and project into a repository.
[ RESEARCH CONTEXT ]
AI literacy is broader than prompt literacy.
These sources motivate the educational direction; they are not evaluations of MLBricks MLB Studio itself.
Students as responsible users and co-creators.
UNESCO’s student framework organizes AI competency across human-centred mindset, ethics, AI techniques/applications and AI system design, with progression through understanding, application and creation.
Read framework ↗AI literacy for life inside and outside the classroom.
The joint framework describes AI literacy as knowledge, skills and attitudes that help learners understand systems, critically evaluate outputs, and use AI ethically and creatively.
Read framework ↗Reduce syntax burden; focus on concepts.
Research on novice visual programming reports that graphical interfaces can reduce syntax-related difficulty and make computational structures more approachable, especially for learners without a computer-science background.
Read study ↗MLBRICKS THESISAI education should be possible before someone decides to become an AI engineer.
A biology student, artist, teacher or physicist should be able to understand the architecture of an AI system, build a small one, evaluate it, and connect it to their own domain. Deeper mathematics and programming can then become tools for going further—not gates that decide who is allowed to begin.
[ START IN A NOTEBOOK ]
Open MLB Studio.
Then change something.
MLB Studio uses the standard notebook/Python environment as its execution host. The visual workspace can run in Jupyter/Kaggle and the full-window interface remains connected to the Python kernel.
PYTHON / JUPYTER
pip install mlbricks-Studio
from mlbstudio import Builder
builder = Builder()
builder.web()[ MLBRICKS STUDIO ]
AI for everyone means everyone should be able to open the system and learn how it works.
Start with a visual model. Bring data from a subject you understand. Train it. Question it. Share it. Then go deeper.