Document Studio
A document studio that reads a request, drafts the reply, and signs the certificate — verbatim values where the source has them, grounded prose where it doesn't, a human in the loop before anything leaves the firm.
Client — Mid-market professional services firm
Sample documents, extracted fields, drafted replies, and signed certificates shown on this page are synthetic — crafted to demonstrate the pipeline, never taken from a client's real paperwork.
Back-office work is paperwork — invoices, agreements, batch records. Document Studio reads them, pulls the fields, drafts the response, and signs the certificate. Same engine, two visual styles: verbatim where the source has the value, grounded where it doesn't.
Triage · Draft · Certificate
Half of inbound is operational — price, stock, order status. The pipeline classifies on arrival and drafts the reply before a rep opens the email.
Inbound · classified on arrival
Could you confirm today's price for the 50mg/ml oral suspension, 100ml bottle, ex-MHRA specials catalog #CMP-0421? We have a prescription for 4 bottles this week.
Catalog lookup
Catalog v6.2 · 412 SKUs
Reply drafted against catalog ground truth — no invented SKUs or prices.
Compliance & send
- · Recipient class verified
- · Jurisdiction · opt-out honored
- · Deliverability · SPF + DKIM pass
- · Message compliance · identity + unsubscribe present
Drafted reply
Interactive / built with synthetic data for this page
[ THE CHALLENGE ]
Professional-services back-office is paperwork on top of paperwork. Operators re-key the same fields out of PDFs into CRMs, draft follow-ups from the same boilerplate, hand-write certificates from the same spec, and re-key all of it into the archive — every handoff is a typo waiting to happen and a half-day waiting to be lost. The team needed every document to land already structured, with the next one already drafted, with nothing invented along the way.
How we built it
From signal to shipped.
Ingest any document
A hybrid parser: deterministic templates for known shapes (invoices, agreements, batch records), OCR + vision for the rest. Auto-detects file type, language, and layout.
Extract the canonical values
A two-layer extractor pulls canonical fields with regex/coord, then a grounded LLM pass handles the messy semi-structured bits. A confidence gate flags anything below threshold for review.
Two-style authoring
The drafted output puts source values in their place, no invention, then writes the prose around them — grounded on the extracted fields or the spec. Reviewers see at a glance which parts are verbatim and which were drafted.
Approve, sign, deliver
Drafts land in review. A qualified reviewer approves outbound, two qualified reviewers sign regulated output, then the workflow routes the draft to the right inbox and the source to the right folder. Every step auditable, end-to-end.
What shipped.
Hybrid ingestion across PDF, DOCX, and image scans — median 3.4s per document.
Two-layer extraction with confidence gating; ~+34% accuracy over a regex-only baseline.
Two-style generation (verbatim vs grounded) makes every output reviewable in seconds.
One engine covers inbound triage, drafted replies, and signed certificates.
[ NEXT PROJECT ]
Sourcing Copilot