Boston Medical Data, a Boston healthcare software company serving anesthesia practices, ambulatory surgery centers and hospitals, runs DocRouter on-premises to process peri-operative document packets for anesthesia revenue cycle management (RCM). DocRouter is an open-source document processing platform from Analytiq Hub, released under the Apache 2.0 licence, which combines OCR, large language models and human review. Its case study, added to the DocRouter site in August 2026, describes the work as a live production deployment.
Packets arrive as mixed fax and PDF bundles that can range from a single page to around 500 pages and often cover many patients and facilities in one file. They mix surgery schedules, cover sheets, demographics, insurance cards, pre-operative forms and anesthesia records. Billing depends on fields such as patient identity, MRN, insurance, procedure timing, providers and ASA status. Staff previously sorted packets by patient by eye and re-keyed the fields into RCM systems, which did not scale with packet volume or page count.
The production path is a native DocRouter workflow triggered by a document.uploaded event for the anesthesia packet tag. It splits each PDF into pages, then runs Amazon Textract OCR and a page-classifier LLM prompt in parallel on every page. The classifier labels each page as a cover, surgery schedule, pre-op form, anesthesia record, demographics sheet, insurance card, ID card or unknown, and pulls name, date of birth and MRN in the same pass. A code step then groups pages into patient sets by MRN or by name plus date of birth, using adjacency heuristics for weakly identified pages, separates schedule and cover pages, builds one merged PDF per patient, and sets a human_review flag when unassigned pages remain. Structured JSON is returned through the API to billing and operations systems.
Extraction runs against a library of schema-driven prompts: a multi-patient surgery schedule extractor covering times, surgeon, MRN, room and procedure across facility formats; patient demographics with primary, secondary and tertiary insurance; insurance cards with member ID, plan, group, Rx BIN/PCN and payer ID; anesthesia records with diagnoses, procedures, surgeons, anesthesia providers and CRNAs, start and stop times, ASA class and block type; a variant for Epic anesthesia records using Epic Procedure Summary care-team rules; and deeper pre-surgery extraction of ASA, PONV risk, airway, medications, labs and history when that documentation is present. Reviewers intervene only on flagged pages and low-confidence edge cases.
The same split, classify, group and extract pattern was first prototyped with Temporal in a separate open-source project, doc-router-temporal, with durable retries and polling; production moved to DocRouter's built-in workflows once they were ready. DocRouter reports that the deployment has processed more than 175,000 documents of 1 to 500 pages each at 99.9% precision on extractions in production, without stating the measurement method or period. Boston Medical Data's own site lists DocRouter founder Andrei Radulescu-Banu as a technology advisor to the company.