Turn any document or messy data source into clean, structured records — with AI doing the heavy lifting and a human making every decision that matters. Four industry modules are built on the same three engines.
Architecture
Every industry module reuses the same engines, the same model gateway and the same human-review discipline. Adding the fourth module needed almost no new machinery — the reuse was the point.
Engine 1 · AI DocumentOS
Documents flow through eight stages. The model extracts and validates; a confidence router decides whether a record goes to a quick check or a detailed review. Either way, nothing reaches the data store without a person accepting it, and every stage writes to an append-only audit log.
Documents arrive by upload or API — PDFs, scans, spreadsheets — and each one gets an audit-log entry from the very first moment.
The model identifies what kind of document this is, which decides the schema it will be read against.
Text is read natively where the file has it, and through OCR where it is a scan.
The model fills the schema's fields from the text, page by page, with confidence for each value.
Rule checks (types, formats, ranges) plus a semantic check catch values that look right but aren't.
High-confidence records go to a fast-lane check; uncertain ones get a detailed review. Confidence picks the lane — never whether a human looks.
A person accepts or rejects every record. Nothing is auto-approved.
Accepted records are written to the data store, with the full trail of who decided what, and when.
Engine 2 · AI Data Agent
Point the agent at sample rows from a messy CSV, spreadsheet or JSON source and it produces a mapping into a clean schema. The mapping is a recipe built only from twelve safe operations — the system's own interpreter carries it out, and anything outside the menu is rejected before it can run.
{ "target_field": "unit_price",
"source_path": "items[0].price",
"ops": [ {"op": "strip"},
{"op": "regex_extract", "pattern": "[0-9.]+"},
{"op": "to_number"} ] }If a model writes Python that then runs on real data, a crafted file or a manipulated model response becomes code running with the system's own permissions — so you need a sandbox around it. Here there is nothing to sandbox: "read a file" or "call the internet" simply can't be expressed in the recipe language.
It is also cheaper. Once a human approves a mapping, a 50,000-row file runs through it deterministically — no model call per row.
A field the agent can't map correctly is left explicitly unmapped rather than shipped as a silent bad guess.
Engine 3 · AI DataOps
A pipeline is generated and tested by the Data Agent, then approved by a person as logic — not re-reviewed on every run.
Each run is compared with the last clean baseline: row-count drift, null-rate drift and schema drift. Plain statistics, not a model's opinion.
Anomalous runs keep their rows in a holding area until a person dismisses or reloads them. Clean runs flow to the warehouse with column-level lineage.
Industry modules
An AI sales agent for WhatsApp and web chat with three paths per message: answer straight away from published content, ask a person for a fast live approval when a reply would commit the brokerage (with a deadline and a safe fallback message), or batch low-stakes CRM updates for a daily skim. Whether a reply is grounded is checked in code, not by asking a second model.
Administrative workflows only — scheduling, prior authorisations, patient messages — reusing RealEstateOS's approval machinery. A conservative classifier flags clinical content, and the Model Gateway refuses flagged input by type, before any model is reached. Every patient-record read is itself logged.
Daily route plans from a deterministic bin-packing optimiser — no model call at all, because the task doesn't need one. Hundreds of stops are reviewed in a fast lane plus per-stop detail, and dispatch happens only when every stop is resolved.
Guest conversations reuse the real-estate routing unchanged. New here: card numbers are scrubbed from messages before they are stored or modelled, and the gateway refuses payment-card data for every task — a second compliance domain on the same refusal mechanism.
Key design decisions
One honest lesson from reuse: the real-estate "viewing request" detector didn't transfer to hotels — guests asking about a "sea view room" would have triggered a property-viewing approval. Routing transferred cleanly; intent detection needed a per-industry switch.
Happy to walk through the architecture and trade-offs.