Extracting data from OSHA logs
Employers keep OSHA's injury log on the official forms or on equivalent layouts from safety software, and summarize it on Form 300A each year. Insurers and contractor prequalification teams collect several years of summaries from every contractor they assess, as signed PDFs or scans.
Underwriters and prequalification reviewers need each case count, the days lost, average employees and total hours worked. Prequalification software turns those into DART and total recordable incidence rates, and scores each contractor against its industry average.
What makes OSHA logs hard
Columns named by letter
Each count sits above a lettered column reference such as (H) or (K), under a multi-line caption that reads differently from the field name.
Two totals, one set of cases
Outcome counts in columns G to J and injury type counts in column M partition the same cases, so their totals must agree.
Example text beside blank lines
The industry description, SIC and NAICS lines print a sample entry beside each answer line, and many employers leave those lines blank.
Parse, extract, verify
Parse
Parse turns the document into reading-order Markdown and a hierarchy of blocks (text, tables, figures), and grounds each block to its exact location on the page.
Extract
You define the fields in a JSON schema, and Extract returns their values in a consistent, structured format, each linked back to where it appears in the Markdown.
Verify
Each value links back to where it appears, down to the character range in the Markdown, and the page and bounding box in the source document, so every result can be verified.
One OSHA log, end to end
About this document: The completed Form 300A summary that the University of Wisconsin-Madison published to meet OSHA's posting requirement. One landscape page, generated from the university's safety records system: case counts, day totals and injury types on the left, establishment and employment details on the right. Page 1 is featured. The form holds aggregate counts only, so no redaction was needed.
Processing objective: The objective is to collect the inputs to a DART incidence rate. Cases with days away and cases with job transfer or restriction, multiplied by 200,000 and divided by hours worked, give the rate. Total days away shows how severe those cases were, and each value grounds to its own box.
Demonstration fields: Cases with days away from work (H), Cases with job transfer or restriction (I), Total days away from work (K), Annual average employees, Total hours worked
The document
1 page
Page 1shown below
Where each value came from
number_of_cases.days_away_from_work
{
"value": 60,
"page": 1,
"range": {
"end": 1177,
"start": 1175
},
"box": {
"xmin": 0.10081,
"ymin": 0.45211,
"xmax": 0.19606,
"ymax": 0.48816
}
}number_of_cases.job_transfer_or_restriction
{
"value": 25,
"page": 1,
"range": {
"end": 1188,
"start": 1186
},
"box": {
"xmin": 0.19592,
"ymin": 0.4521,
"xmax": 0.31125,
"ymax": 0.48817
}
}number_of_days.days_away_from_work
{
"value": 938,
"page": 1,
"range": {
"end": 1454,
"start": 1451
},
"box": {
"xmin": 0.02144,
"ymin": 0.60635,
"xmax": 0.13055,
"ymax": 0.65413
}
}employment.annual_average_employees
{
"value": 34321,
"page": 1,
"range": {
"end": 3374,
"start": 3368
},
"box": {
"xmin": 0.79044,
"ymin": 0.59046,
"xmax": 0.94552,
"ymax": 0.63317
}
}employment.total_hours_worked
{
"value": 46120878,
"page": 1,
"range": {
"end": 3457,
"start": 3447
},
"box": {
"xmin": 0.79042,
"ymin": 0.63287,
"xmax": 0.9454,
"ymax": 0.66247
}
}How to try it yourself
Everything on this page is reproducible. The sample, its schema, and the Parse and Extract API outputs are published. You can repeat the same two calls from whichever surface you already build on.
- Claude Skilldescribe the job; the agent writes it
- CLIno code at all
- MCPexpose ADE docs as an agent tool
- REST APIany language
- Pythonlandingai-ade
- TypeScriptlandingai-ade
Documents into data
Accurate
The best accuracy on the difficult documents other tools get wrong, such as messy scans, complicated tables, checkboxes and handwriting.
Auditable
Every answer cites its sources. Values come back with an audit trail (page, bounding box, character range) so a human can review them.
Agentic
Applies multi-step reasoning, and plans before it acts. Handles documents it has never seen before and adapts to variations.
API-first
Composable APIs that drop into any tech stack and can be called from anywhere. Easy to build on and embed in your pipelines and products.
Affordable
Cost-competitive rates with complete price transparency. Spend controls before you run, and every job reports its exact cost.
Enterprise-grade
Built for production and scale from day one. Fully observable and monitored, running securely in our cloud or yours.
Built for high-volume, high-throughput production
- Zero data retention
- Priority rate limits
- Enhanced support
- VPC and on-prem deployment
Available on Team and Enterprise plans. VPC and on-prem deployment require an Enterprise plan.
