Document AI · Experimentation → Production

Find the best Document AI workflow for your documents.

DocOS is an experimentation-to-production platform for document classification and key information extraction. Upload labeled documents, test different AI workflow designs, compare performance, and deploy the winning pipeline as an API.

Request Early AccessSee Example Workflow →

Built for teams working on invoices, forms, claims, contracts, financial documents, and other high-volume business documents.

invoice-extraction · candidate C
running evalDeploy ↗
Document Input5,000 PDFsOCRtesseractLayout Parserlayout-v3Embeddingtext-embedViT Classifiervit-b/16LLM Extractorllm-xlValidationrules + schemaHuman ReviewfallbackDeploy · APIv1 endpointF1 0.917 · acc 94.8%
Compose from
OCRLayout parserEmbeddingsViTTraditional MLLLM extractorsValidation rulesHuman review
The workbench

A visual workbench for Document AI experiments.

Drag components onto the canvas, wire them into candidate workflows, and watch evaluation update on your labeled data — no notebooks to babysit.

Candidate CCandidate BCandidate A+
InputPDFOCRtesseractViTvit-b/16LLM Extractorllm-xlValidationschemaDeployv1
The problem

Building Document AI workflows is still too manual.

Most teams test document workflows across notebooks, scripts, spreadsheets, and disconnected dashboards. Comparing OCR tools, ML models, embeddings, LLMs, prompts, validation rules, and deployment options is slow and difficult to reproduce.

01

Too many workflow choices

OCR or native text? Traditional ML, embeddings, ViT, or an LLM? Rules or prompts for extraction? Every combination is a new experiment.

OCRno-OCRMLViTembedLLMrulesprompthybrid
02

Hard to compare results

Accuracy, latency, cost, and failure cases end up scattered across notebooks, spreadsheets, and one-off dashboards. Nothing is reproducible.

notebook_v3.ipynbresults_final2.xlsxeval_dash.png
03

Hard to productionize

The winning experiment still needs glue code, versioning, and monitoring before it can become a reliable endpoint your systems can call.

experiment.py→glue code→versioning→?
How it works

From labeled documents to production API.

One connected loop — no exporting between tools, no rebuilding the winner from scratch.

  1. Step 1

    Upload your dataset

    Import ground-truth documents and split them into train, validation, and test sets.

  2. Step 2

    Build workflow candidates

    Drag and drop OCR, classification, extraction, validation, and model components.

  3. Step 3

    Compare performance

    Evaluate accuracy, latency, cost, confidence, and failure cases side by side.

  4. Step 4

    Deploy the winner

    Turn the best workflow into a versioned production endpoint your systems can call.

AI-native

Let an agent build the workflow — live on the canvas.

Connect any MCP-capable agent. It edits workflows through safe, scoped tools, and every change is reflected in the builder in real time.

connected · MCPdocos-workflows
New task: classify invoices and pull vendor, total, and tax. I'll start a fresh workflow.
agent ▸create_workflow(args)✓
{
  "name": "invoice-extraction",
  "task": ["classify", "extract"],
  "fields": ["vendor", "total", "tax", "date"]
}
→ workflow created
workflow.json · 0 nodes
invoice-extractionbuilding…

Illustrative — an agent reasoning and editing the workflow through MCP tools; the builder updates as each call lands.

Worked example

Example: invoice classification and extraction.

One dataset, four candidate designs, one evaluation surface.

Input
Dataset5,000 labeled invoice PDFs
Ground-truth fields
vendorinvoice_nodatetotaltaxline_items
Target taskClassify invoice type · extract fields
Workflow candidates
  • OCR + rules-based extraction86.2%baseline
  • OCR + embedding classifier + LLM extractor94.8%★best accuracy
  • Layout parser + ViT classifier + field extractor92.1%best balance
  • Native PDF text + LLM extractor + validation rules90.4%lowest cost
Best workflowOCR + embeddings + LLM
Accuracy / cost / latency94.8% · $18.60 / 1K · 3.8s
Failure cases42 flagged for review
EndpointPOST /v1/…/invoice-extraction
Evaluation

Compare workflows by accuracy, cost, latency, and failure cases.

Every candidate is scored on the same held-out test split — so the comparison is apples to apples.

WorkflowAccuracyField F1LatencyCost / 1KStatus
OCR + Rules86.2%81.5%1.2s$4.20Baseline
OCR + Embeddings + LLM94.8%91.7%3.8s$18.60Best Accuracy
Layout + ViT + Extractor92.1%89.3%2.5s$11.40Best Balance
Failed examples · winning workflow42 of 750 test docs
  • invoice_3182.pdftotalhandwritten amountconf 0.41
  • invoice_0947.pdftaxmulti-currency lineconf 0.52
  • invoice_4410.pdfvendorlogo-only headerconf 0.38
Deployment

Deploy the winning workflow without rebuilding it from scratch.

Once a workflow wins, DocOS is designed to turn it into a production endpoint that integrates into your existing business systems.

REST API endpoint
Versioned workflows
Test & production environments
Monitoring
Rollback
Webhook support
SDK-ready integration
Human-review fallback

Prototype flow — deployment behavior shown is illustrative.

endpointproduction
# run the deployed workflow
POST /v1/workflows/invoice-extraction/run
# → 200 OK
{
  "document_type": "invoice",
  "fields": {
    "vendor": "Acme Inc.",
    "invoice_number": "INV-2048",
    "total": "$4,210.00"
  },
  "confidence": 0.94
}
Use cases

Built for high-volume business documents.

DocOS focuses on document classification and key information extraction.

Invoice classification & extraction

Route by type, pull vendor, totals, tax, and line items.

classification · KIE

Insurance claims intake

Classify claim forms and extract structured fields at intake.

classification · KIE

Bank statement processing

Parse transactions and balances from mixed-layout statements.

KIE

Contract routing & clause extraction

Sort contracts and surface the clauses that matter.

classification · KIE

Healthcare administrative forms

Extract fields from intake and coverage documents.

KIE

Logistics & bills of lading

Read shipment documents into structured records.

KIE

Tax & financial documents

Classify and extract from high-volume financial paperwork.

classification · KIE
Why DocOS

One platform for the full Document AI experiment loop.

Not just OCR. Not just labeling. Not just model hosting. DocOS connects dataset → experiment → evaluation → comparison → deployment.

Current approach
The DocOS approach
Test models manually in notebooks
Compose workflows visually
Compare results in spreadsheets
Evaluate pipelines in one dashboard
Optimize only for accuracy
Compare accuracy, cost, latency, and failure cases
Rebuild experiments for production
Deploy the winning workflow as an endpoint
Scripted, one-off automation
Agent-drivable via MCP, reflected in the UI
Scattered tools
Dataset-to-deployment in one workflow
The bigger picture

Starting with experimentation. Building toward the operating system for document workflows.

DocOS begins with experimentation-to-production for document classification and key information extraction. Over time, DocOS aims to support the full lifecycle of AI workflows involving documents. And because DocOS is AI-native, every capability is exposed to agents through MCP.

Ingestion
Labeling
Experimentation
Evaluation
Deployment
Monitoring
Improvement

■ available in the first product■ on the roadmap

Early access

Ready to benchmark your Document AI workflow?

Join early access and see how DocOS can help your team discover, evaluate, and deploy the best workflow for your documents.

  • Bring your own labeled documents and models
  • Compare candidates on your held-out test split
  • Deploy the winner as a versioned API

No spam. We’ll only email you about DocOS early access.