---

Marie-AI and M3 Forge are two parts of the same platform direction. Marie-AI runs the backend
document processing work. M3 Forge gives teams a way to configure that work, review exceptions,
monitor execution, and operate the system without living inside scripts and logs.

Marie-AI runs the work. M3 Forge gives people a way to design, review, monitor, and operate the work.

When we say “documents,” we do not only mean PDFs. The platform is intended for a wide range of
document and file types, including scanned images, digital documents, forms, invoices, EOBs,
claims, records, attachments, and other structured, semi-structured or unstructured content.

Why this exists

A lot of document AI projects start as a prototype. You take a sample document, run OCR or send it to a model,
get back some JSON, and prove that extraction is possible.

That is useful, but it is not the hard part.

The hard part is turning that prototype into a system that can run every day:

• Large files and multi-page document packages need to be split, partitioned, or routed intelligently.
• Pages need to be classified before the right extractor runs.
• Different document types need different processors.
• Fields need validation and business rules.
• Uncertain results need human review.
• Failed jobs need safe retries.
• Operators need visibility into throughput, failures, confidence, and review quality.

Marie-AI and M3 Forge are built for that production layer. The goal is not just to extract data
once. The goal is to create a repeatable workflow that can be operated, reviewed, improved, and
trusted.

What Marie-AI does

Marie-AI is the runtime and execution layer. It is responsible for the backend work involved in
document processing and AI workflow execution.

That includes:

• OCR and document understanding for scanned images, digital documents, PDFs, forms, and other document formats.
• Document and page classification so the right processing path can be selected.
• Extraction of fields, tables, and structured business data.
• Workflow execution for multi-step document pipelines.
• Job orchestration across workers, queues, processors, and runtime services.
• Model integration with OCR engines, ML models, VLMs, LLMs, and custom processors.

Marie-AI is the part of the system that actually processes documents, images, scans, and structured/unstructured files through the workflow.

What M3 Forge adds

A runtime alone is not enough once a team has multiple workflows, processors, reviewers, document
types, and deployment environments.

M3 Forge provides the product and control plane around the runtime.

It adds:

• Workflow design for defining how documents move through classification, extraction, validation, and review.
• Processor management for organizing reusable document AI components.
• Human review for low-confidence or exception cases.
• Monitoring for runs, failures, throughput, traces, and operational health.
• Deployment control for cloud, hybrid, and self-hosted operating models.

M3 Forge exists so the people building and operating workflows do not have to work only through
backend code, command-line tools, and logs.

How a production workflow comes together

Consider a high-volume document processing workflow that receives mixed files from customers, vendors, internal teams, or automated systems.

A workflow might send a document, image, scan, attachment, or mixed batch through a model-assisted extraction step and ask for structured output. That may be one useful step, but
a production workflow needs more structure, controls, and visibility around the entire process.

A production process may need to:

1. Receive a document, image, scan, email attachment, or batch of mixed files.
2. Split large documents into manageable sections.
3. Classify pages or document sections.
4. Run OCR and layout analysis.
5. Extract fields, tables, sections, entities, and other structured business data.
6. Validate required fields and business rules.
7. Route uncertain results to human review.
8. Store corrections and final structured output.
9. Monitor failures, throughput, and review quality over time.

Marie-AI executes the processing steps. M3 Forge gives users a way to configure, observe, review,
and operate the workflow.

Why not just use an LLM?

LLMs are useful, but they do not replace the rest of the system.

A production document AI workflow still needs OCR, layout handling, page classification, validation,
review queues, retries, versioning, monitoring, deployment controls, and integration with downstream
systems.

The model may be one part of the workflow. It should not be the entire workflow.

Where this is useful

This platform direction is useful when document processing has real operational requirements:

• High-volume business documents
• Scanned images and digital forms
• Invoices, remittance files, and operational records
• Applications, contracts, packets, and case files
• Mixed-format batches from email, portals, or internal systems
• Large multi-page and mixed-format document processing
• Workflows that require validation, exception handling, or human review

The common pattern is that these documents cannot be treated as simple text. They have layout,
structure, business rules, exceptions, and downstream consequences.

Deployment options

Different teams have different infrastructure and data requirements. Some want a managed platform.
Some need the runtime close to their data. Some want to self-host everything.

• Cloud: the platform is operated as a managed service.
• Hybrid: M3 Forge provides the control plane while Marie-AI runs in the customer environment.
• Self-hosted: both M3 Forge and Marie-AI run inside the customer environment.

That flexibility matters for organizations that process sensitive documents or need stronger control
over infrastructure and data movement.

Current focus

The current focus is building the practical pieces needed for production document AI:

• Make workflows easier to define and operate.
• Make processors easier to manage and reuse.
• Make human review part of the workflow instead of an afterthought.
• Make runtime behavior observable.
• Make cloud, hybrid, and self-hosted deployments realistic.

A single extraction step can prove that useful data can be captured. A platform helps a team run that extraction
process consistently every day.

Learn more

• Marie-AI on GitHub
• M3 Forge website
• M3 Forge documentation
• Self-hosted deployment notes