§ About us
Documents should explain themselves.
STELE exists because the world's most important information still arrives as PDFs, scans, and photographs — and the systems built to read them were designed for a cloud era that priced data movement like a toll road. We rebuilt the whole discipline on the edge.
Why we started
Every finance, legal, and operations team we have ever worked with runs on documents nobody can query. Invoices are re-keyed by hand. Contracts are “searched” by opening them one at a time. Audits are archaeology. The industry’s answer — template-driven OCR hosted on hyperscalers — solved 2015’s problem and then charged rent on it: egress fees to move your own files, professional services to repair brittle templates, and AI black boxes with no record of what was extracted, by what, or why.
STELE started from a different premise: the document is the database. If a machine can read a document the way an analyst does — noticing what kind of document it is, pulling out the fields that matter, admitting when it is not sure, and citing its sources — then the warehouse of PDFs your organisation sits on becomes structured, searchable, forecastable data. That machine should run where the data already lives, cost cents rather than dollars, and write down everything it does.
Why “STELE”
A stele is a stone slab raised to make a record public and permanent — laws, treaties, ledgers. The Rosetta Stone is a stele. It is the oldest technology humanity has for turning a document into something durable, verifiable, and readable by anyone who comes after. That is the standard we hold extraction to: not just fast answers, but answers that stand as a record.
The principles we build by
1. Data gravity is a tax — refuse to pay it
Your documents are stored, parsed, embedded, and queried inside the same edge network. Nothing crosses a metered boundary, which is why our per-page economics undercut legacy IDP by roughly 40% — and why your data never tours the public internet between services.
2. Confidence must be a number, not a vibe
Every extraction ships with a confidence score, passes an arithmetic validation gate, and is checked against your corpus for statistical anomalies. An answer without a stated uncertainty is marketing, not engineering.
3. Humans stay in the loop — by design, not apology
When the system is unsure, it says so and routes the document to a review inbox instead of shipping a guess. Your corrections are recorded in the audit ledger and become few-shot memory that improves the next extraction. The model works for your reviewers, never around them.
4. The audit trail is a feature, not a compliance chore
Every ingestion, extraction, escalation, and human correction is appended to an immutable ledger. If a regulator, auditor, or court asks “where did this number come from?”, the answer is one query away.
How we build
STELE runs entirely on Cloudflare’s developer platform — Workers for compute, R2 for object storage, D1 for the ledger, Queues and Workflows for durable orchestration, Vectorize for the semantic index, and Workers AI for open-weight language models hosted in the same network. There are no servers to patch, no regions to replicate, and no third-party AI API in the data path: documents are processed by models running inside the same infrastructure that stores them, and customer content is never used to train models.
We publish what the pipeline actually emits — the JSON on our landing page is the real extraction shape, not a mock-up — and we document our privacy practices, terms, and data-processing commitments in plain language.
STELE today
| 10 document types detected from content — invoices, receipts, contracts, purchase orders, bank statements, resumes, and more — with a self-describing schema for everything else | 768-dimension semantic index built at ingest, powering search and cited chat over the whole corpus |
| 100% of processing on the edge network where your documents are stored — zero egress, zero external AI APIs | Every extraction validated, scored, and recorded in an append-only audit ledger |
Talk to us
We are a small, senior team of infrastructure and machine-learning engineers, and we answer our own email.
- Sales and partnerships: sales@stele.example
- Security reports: security@stele.example (we acknowledge within one business day)
- Privacy and data protection: privacy@stele.example
Or skip the email: create a workspace and process your first hundred pages free, every month.