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
Local-first Document AI · Experiment → Compare
Find the best Document AI workflow for your documents.
Drop in receipts, compare local OCR + Haiku with direct vision + Sonnet, and inspect field quality, cost, confidence, and failures. Your project, files, and local OCR stay in the browser; LLM inputs go directly to Anthropic.
No signup · Includes 20 real CORD receipts · Hosted deployment is on the roadmap.
Seeded preview for orientation — not a recorded model execution.
Workflow palette
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.
Accurate · visionFast · local OCR+
The problem
Building Document AI workflows is still too manual.
Every combination of OCR, models, prompts, and rules is a new experiment — and the results end up scattered, slow to compare, and hard to reproduce.
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 real receipts to a measured workflow.
Four steps, entirely in your browser — from a real sample to a comparison you can trust.
- Step 1
Open the receipt sample
Start with 20 real CORD receipt photos and labels, or drag in your own documents.
- Step 2
Build workflow candidates
Wire local OCR or direct vision into an LLM extractor, validation, and output.
- Step 3
Run and compare
Use your Anthropic key to measure field F1, exact match, latency, cost, and confidence.
- Step 4
Inspect and choose
Open mismatches and traces, adjust the workflow, then choose the trade-off that fits.
Worked example
Example: extract totals from real receipt photos.
One 20-image CORD sample, two runnable candidates, one comparison surface.
Input
Dataset20 labeled CORD v2 receipt images
Ground-truth fields
Target taskExtract three money fields across varied receipt formats
Workflow candidates
- Fast · local OCR + Haiku
- Accurate · vision + Sonnet
Seeded comparison leadAccurate · vision + Sonnet
Field F1 / exact match92.3%* · 83.3%*
Cost / 1K · latency p50$5.65* · 2.5s*
Next stepRun it with your Anthropic key
* Illustrative, precomputed sample values for orientation — not recorded Anthropic executions.
Evaluation
Compare workflows by accuracy, cost, latency, and failure cases.
The console runs every candidate on the same split. This landing-page preview uses seeded, precomputed values so you can see the shape of the comparison before running it yourself.
| Workflow | Field F1 | Exact match | Latency p50 | Cost / 1K | Status |
|---|---|---|---|---|---|
| Fast · local OCR + Haiku | 76.9%* | 50.0%* | 1.6s* | $0.50* | Lower cost |
| Accurate · vision + Sonnet | 92.3%* | 83.3%* | 2.5s* | $5.65* | Higher field F1 |
Seeded failure example · accurate candidate1 illustrative mismatch
- cord_receipt_009.jpgtotalseeded one-digit mismatchconf 0.74
* Illustrative, precomputed sample values — not recorded Anthropic executions.
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 · KIEInsurance claims intake
Classify claim forms and extract structured fields at intake.
classification · KIEBank statement processing
Parse transactions and balances from mixed-layout statements.
KIEContract routing & clause extraction
Sort contracts and surface the clauses that matter.
classification · KIEHealthcare administrative forms
Extract fields from intake and coverage documents.
KIELogistics & bills of lading
Read shipment documents into structured records.
KIETax & financial documents
Classify and extract from high-volume financial paperwork.
classification · KIEWhy DocOS
One browser workspace for the Document AI experiment loop.
DocOS connects dataset → candidate → evaluation → comparison in one place, instead of a pipeline stitched together by hand.
Current approach
◈ The DocOS approach
Test models manually in notebooks
Compose runnable workflows visually
Compare results in spreadsheets
Evaluate candidates in one local dashboard
Optimize only for accuracy
Compare field quality, cost, latency, confidence, and failures
Upload every document to a new service
Keep files and local OCR in the browser
Build a demo dataset from scratch
Start with 20 labeled, real CORD receipts
Lose the details behind one score
Inspect normalized field matches and per-node traces
Product direction · illustrative
A future agent could build the workflow on the canvas.
This is the planned hosted path: scoped MCP tools for workflow editing, evaluation, and deployment. MCP control is not available in the local prototype today.
concept · MCPfuture · docos-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
Illustrative concept — not part of the local console today.
What’s next
Today: local experiments. Next: hosted operation.
The prototype covers the experiment loop today. The hosted product will turn a chosen workflow into a managed, monitored endpoint your systems can call.
Ingestion
Labeling
Experimentation
Evaluation
Deployment
Monitoring
Improvement
■ available in the local prototype■ on the roadmap
The hosted layer · paid roadmap
REST API endpoint
Versioned workflows
Test & production environments
Monitoring
Rollback
Webhook support
SDK-ready integration
Human-review fallback
Roadmap concept — the local prototype does not create a production endpoint.
# planned hosted API POST /v1/workflows/receipt-extraction/run # illustrative response { "document_type": "receipt", "fields": { "subtotal": "259000", "tax": "52416", "total": "302016" }, "confidence": 0.94 }
Early access
Want the hosted version of DocOS?
Join early access if your team needs managed deployment, monitoring, collaboration, or higher-volume evaluation.
- Explore local OCR versus direct vision on real receipts
- Run candidates with your own Anthropic key
- Help shape hosted deployment and team workflows
Prefer to explore first? No signup required. Open the free console →