Use case · Data extraction
Ten thousand invoices.
One clean table.
The quarter's vendor invoices arrive as scans — crooked, stamped, occasionally handwritten. Sixteen workers read them all before the second coffee, and confess to the ones they can't.
Nobody's judgment improves on invoice four thousand. Humans get slower and less careful as the pile grows. Sixteen workers read page one and page ten thousand with exactly the same attention — and admit it when they aren't sure.
7:02 AM
The folder nobody wanted to open.
Quarter close. A finance analyst has 10,412 scanned PDFs and a warehouse table that needs them as rows. Instead of a data-entry vendor and a three-week wait, one API call with a schema attached. Totals come back in integer minor units — no floating-point money, ever.
POST /api/v1/swarms{ "objective": "Extract line items from Q2 vendor invoices", "inputArtifact": "invoices-q2.zip", // 10,412 scanned PDFs "workers": 16, "outputSchema": { "vendor": "string", "invoiceNumber": "string", "issuedAt": "date", "totalMinorUnits": "int", "currency": "string" }, "aggregatorTask": "Validate rows against the schema, dedupe, emit one table + an exceptions queue", "budgetUsd": 120.00}7:03 AM
Sixteen readers, none of them bored.
The archive shards across sixteen sandboxed workers — roughly 650 invoices each, all reading at once. A worker on its 600th invoice is exactly as careful as it was on its first.
Every row carries a confidence score. Below the threshold, a worker doesn't guess — it routes the document to an exceptions queue and moves on. The $120 ceiling is hard; the run can never cost more than the number in the request.
10,000 PDFs → one clean table
7:58 AM
One table, and an honest list of doubts.
Fifty-six minutes in, the aggregator hands back a validated, deduplicated table: 10,372 rows that conform to the schema, every one traceable to the worker and the page it came from.
And a queue of 40 it refused to guess on — handwritten totals, a stamp over the amount, a suspected duplicate. Each with the reason it was held, not just a blank.
exceptions queue · 40 of 10,412
9:15 AM
The audit trail wrote itself.
The analyst clears the 40 exceptions by hand — the 0.4% that genuinely needed a human — and the table ships to the warehouse before the 10 AM close meeting.
Beside it, the ledger: the hold, the metered GPU-seconds, the unused budget returned — append-only, line by line. When audit asks how a scanned stamp became a row, the answer is a job id and a receipt, not an intern's memory.
the ledger never lies — append-only
The data-entry vendor quoted $0.19 per invoice and three weeks — about $1,978. This ran 56 minutes and cost $84.10, and it told you exactly which 40 invoices it wasn't sure about. Vendors don't do that.
0
workers reading in parallel
0 min
10,412 scans to a validated table
0
exceptions routed to a human — 0.4%
The same shape, everywhere
Anywhere messy documents hide a clean table, a swarm wins.
Contract abstraction
Renewal dates, caps, and termination clauses pulled from a thousand PDFs into one register.
Claims intake
Forms, photos, and adjuster notes structured into a case record the moment they arrive.
Resume screening
Every application mapped to the same rubric — the 400th read as fairly as the 1st.
Catalog normalization
Supplier spreadsheets in nine formats reconciled into one schema, with a diff per vendor.
Lab and field reports
Handwritten readings and scanned tables digitized with confidence scores on every cell.
Email-to-order entry
Purchase orders buried in inbox threads extracted, validated, and queued for approval.
Keep exploring
Give your agent a workforce.
Start free. Spawn your first swarm in minutes. Pay only for the GPU-seconds you use.