We engineer production LLM pipelines that extract, generate, validate, and search across high-volume document sets - bond official statements, financial reports, scanned filings, regulatory disclosures. Two named US municipal finance production references and a multi-year US healthcare-tech deployment. Built for regulated environments and ready for FDTA 2027 by design.
Four stages, one pipeline. The same architecture handles a 247-page bond official statement, a scanned insurance form, or a regulatory filing - the inputs vary, the flow does not.
Bond statements, regulatory filings, scanned forms, payor contracts. 200 to 1,000+ pages each, hundreds per month.
OCR with layout preservation. Documents classified by type, key fields identified, scanned exhibits handled.
LLM pipeline pulls 100+ structured fields. Cross-validates against business rules and MSRB / SEC schemas.
Structured data shipped to your CRM, treasury, or filing repository. Audit trail logged. FDTA-ready manifest attached.
Kyotu Technology is an AI engineering company based in Warsaw, Wroclaw (Poland, European Union), and Austin (Texas, US), with local presence in Los Angeles, Las Vegas, and Pennsylvania (US). We design, build, and operate production document intelligence systems - LLM-powered extraction pipelines, automated POS/FOS generation, multi-source data aggregation, and validation layers - for clients in US municipal finance, banking, healthcare, and high-volume operational environments. ISO 9001 / ISO 27001 active. DUNS 679252803.
Organizations processing high-volume regulated documents - municipal bond official statements, financial filings, healthcare insurance plans, government records - hit the same wall: senior analysts spending hours reading 200-1,000 page PDFs, manually copying values into spreadsheets, and producing inconsistent outputs under time pressure. DocMind AI replaces that loop with production LLM pipelines.
Each bond official statement runs 200 to 1,000 pages with 100+ data points buried in tables, footnotes, and dense legal language. Manual extraction by an analyst takes 4 to 8 hours per document.
→ Bottleneck. Backlog. Cannot scale headcount fast enough.
Manual data entry introduces inconsistencies that propagate into pricing models, compliance reports, regulatory filings, and audit trails - long after the analyst has moved on.
→ Rework. Compliance findings. Re-issued reports.
The US Financial Data Transparency Act requires continuing disclosures in machine-readable format by 2027. Every municipal advisor, broker-dealer, and underwriter is now under regulatory deadline to digitize their document pipeline.
→ Industry-wide deadline. Retrofit cost. Buy or build decision.
Most document AI vendors do one thing: pull data out of PDFs. DocMind does both - extract data from existing documents and generate new ones from source data, templates, and historical references. The same client deploying both tracks gets full lifecycle automation in a single pipeline.
We ingest 200 to 1,000 page PDFs - bond official statements, prospectuses, regulatory filings, scanned forms - and produce structured data validated against business rules. This is the path live at the Municipal Advisory Council of Texas.
OCR with layout normalization, table recognition, embedded chart parsing. Handles text-based, scanned, and mixed-format collections.
Schema-driven LLM extraction of 100+ parameters per document. Custom prompt engineering per document type and client schema.
Business-rule validation, cross-reference consistency checks, derived value calculation. Confidence-thresholded fallback boosts 95% to 99.5%.
Structured data via REST API, direct database write, or flat-file export. Full audit trail for regulated environments.
We turn source data, historical templates, and regulatory rules into draft Preliminary and Final Official Statements (POS / FOS) ready for legal review. This is the path live at SAMCO Capital Markets, with FDTA-ready structured output by design.
Pull issuer financial data and disclosure history from MAC of Texas, EMMA, State Comptroller, and historical document corpus. Excel Data Prep Agent assembles the master workbook.
Smart Source Picker selects the most relevant historical POS/FOS as a drafting reference using retrieval over the firm corpus. LLM composes the new document section by section with traceable citations.
Output checked against MSRB rules, SEC Rule 15c2-12, and firm-specific drafting standards. Missing fields panel routes the analyst directly to gaps that need human review.
Word document plus structured-data manifest, machine-readable for FDTA 2027 compliance from day one. Audit trail covers every prompt, source citation, and analyst override.
The Financial Data Transparency Act mandates that continuing disclosures from US municipal issuers, broker-dealers, and underwriters must be filed in machine-readable format by 2027. Every municipal advisor and underwriter is now under a regulatory deadline to digitize their document pipeline.
DocMind AI produces structured, machine-readable output as a primary artifact - not as an afterthought. Both extraction and generation tracks emit FDTA-aligned data alongside the human-readable document.
Every generated POS or FOS ships with a paired data manifest. Every extracted document outputs a typed schema. Both formats target FDTA structured-data requirements.
Every prompt, source citation, validation rule fired, and analyst override is logged. Compliance and audit teams get a defensible record of how each document was produced.
DocMind already pulls from EMMA, MAC of Texas, State Comptroller, and similar issuer-data sources. The same integration layer feeds FDTA-format outputs back to filing systems.
Vendors that bolt on FDTA compliance in 2026 will ship inconsistent, low-confidence outputs. DocMind treats structured data as the primary artifact from day one of every deployment.
Document AI in municipal finance fails when the vendor does not understand the underlying instruments, regulators, and data sources. DocMind is built around named instruments and named workflows. The list below is operational vocabulary on every project, not marketing copy.
Pre-pricing disclosure document for a municipal bond offering. 30 to 100 pages of dense legal and financial language.
Post-pricing finalized disclosure, filed with EMMA within 7 business days of pricing. Locked legal record.
Companion documents to a bond issuance. Each has its own data structure, each parsed by DocMind in production.
Permanent School Fund Guarantee, Municipal Utility Districts, Tax Increment Reinvestment Zones, Public Improvement Districts. Domain instruments most generic OCR vendors do not even know exist.
~5,000 issuer-level financial summaries maintained by MAC of Texas. Daily-updated source data for every Texas issuer.
Federal continuing-disclosure rule for municipal securities. The validation backbone of every FOS we generate or extract.
Fair-dealing rule and municipal advisor duty rule. Constrain how much DocMind can automate and where the analyst must stay in the loop.
Federal mandate for machine-readable disclosures by 2027. The strongest regulatory tailwind for digitizing the document workflow.
Every municipal advisor on the buy side is MSRB-registered. We design DocMind workflows that fit the licensed-advisor compliance posture.
For the healthcare-tech track, all data handling is HIPAA-aware. Per-project compliance scoping, not blanket claim.
Official State Information Depository for Texas since 1995. ~5,000 TMRs. The primary data source for any Texas-issuer workflow.
MSRB's national repository for municipal disclosure documents. Mandatory filing destination for every issuance.
State and county-level financial data. 254 appraisal districts in Texas with no unified API - we know which to integrate and which to leave to humans.
For the healthcare-tech track: Onederful, ChangeHealthcare, OpenDental, NexHealth. Direct integrations for insurance plan data aggregation.
Adjacent and competitive tooling. We understand the workflow our clients already have and where DocMind fits next to it.
Three primary verticals where we have named production references. Six adjacent verticals where the same DocMind patterns apply directly. We do not claim domains we have not shipped to - hover any card for the proof line.
POS / FOS extraction and generation, Notice of Sale, Bond Resolution, Remarketing Memorandum, continuing disclosures. SEC Rule 15c2-12 compliance data. Texas-specific structures (PSF, MUD, TIRZ, PID).
Multi-source data aggregation across insurance carrier portals and APIs, payor schema normalization, plan verification automation. HIPAA-aware architecture.
Prospectus and offering memorandum data extraction, pricing data, covenant terms, financial ratios from POS / FOS / Series documents. Pre-IPO and underwriting workflow acceleration.
High-volume financial document processing: loan docs, AML / KYC packets, financial statements, due-diligence materials. Direct architectural fit, target vertical.
Administrative document automation: applications, permits, identification documents, tax filings, agency disclosures. Audit trail and FOIA-ready output. Target vertical.
Contract term extraction, obligation tracking, due-diligence automation, party / date / clause indexing. Direct architectural fit, target vertical.
Policy ingestion at volume, claims-data extraction, regulatory filings, underwriting input automation. Different from healthcare-tech (carrier-side, not provider-side). Target vertical.
Multi-million-document archives, recurring operational document flows, scanned-image collections, OCR-heavy environments. Cost-per-document optimization is the primary value driver.
Procurement docs, supplier contracts, internal documentation at multi-million-document scale, RAG-aware document search. Cross-vertical pattern.
Three production deployments with verifiable metrics. Two are publicly attributed - Municipal Advisory Council of Texas and SAMCO Capital Markets, both in Austin, Texas. The third is anonymized at client request and flagged as adjacent capability. We do not claim cases we have not shipped.
The Municipal Advisory Council of Texas (MAC) - the only comprehensive financial data source for Texas municipal bonds since 1954, and the official State Information Depository for Texas since 1995 - needed to automate processing of thousands of municipal bond documents per month. Before AI, every document required manual analyst review. DocMind now handles Final Official Statements, Bond Resolution / Orders, Notices of Sale, Preliminary Official Statements, and Remarketing Memoranda - each with hundreds of data points critical for SEC Rule 15c2-12 compliance. Nationwide expansion to all US states is the active next step.
Since late August, our team has been using AI to streamline the way we handle and analyze key financial documents. The system has successfully parsed 5,236 documents through the API, with all processed data captured and stored for further use. Laura Slaughter / Executive Director, Municipal Advisory Council of Texas
SAMCO Capital Markets - a 100% employee-owned boutique investment bank headquartered in Austin, Texas, with offices across Texas, Missouri, and Kansas - engaged Kyotu Technology to automate Preliminary and Final Official Statement drafting for municipal bond issues. DocMind pulls source data from MAC of Texas, EMMA, and the State Comptroller, composes a draft document section-by-section, and validates against MSRB and SEC rules. The output is structured-data ready for FDTA 2027 from day one.
Anonymized US healthcare-tech ScaleUp client (Chicago, Illinois). Long-running engagement since October 2021. The system aggregates dental insurance plan data across multiple insurance carrier portals and APIs (including Onederful and ChangeHealthcare integrations), normalizes data into a unified schema via custom rules and scoring engines, and replaces a 20-minute manual phone-call verification process with a sub-minute automated flow. HIPAA-aware architecture. We list this as proof-of-pattern diversity, not as a strict document-extraction case.
These are the technologies live in our shipped DocMind deployments. Items marked “in development” are under active build for SAMCO Capital Markets and not yet in production at any client. We do not list capabilities we have not deployed.
Currently shipping for SAMCO Capital Markets. Not yet in production at any client. We will mark this category “active” once deployed.
We do not pitch with slides and stay until requirements are signed. We start with your actual documents, build a measurable pilot, and only then commit to production. Each phase has a defined output - not a demo, not a deck.
Share 5-10 sample documents under NDA. We analyze document types, identify extraction parameters, and return a feasibility report with confidence ranges per parameter.
48h / freeWe build a working extraction or generation pipeline for one document type with a sample dataset. You see real precision metrics before committing to full deployment.
2-4 weeksCross-validation against ground-truth dataset. Fallback mechanism design. Confidence-threshold calibration. Missing-fields panel set up for analyst review workflow.
In-pilotFull pipeline deployment, system integration, monitoring, audit-trail logging. SLA-backed for enterprise deployments. On-prem or private-cloud option for regulated environments.
4-8 weeksOngoing maintenance, model updates as LLM providers evolve, throughput scaling as document volume grows, audit support, FDTA-format migration. Optional managed-service.
ContinuousHow a production DocMind AI deployment compares against manual analyst processing and against generic OCR / off-the-shelf document tools. Domain-specific tuning is what separates 95% precision from 99.5% with fallback.
| Dimension | Manual processing | Generic OCR / off-the-shelf | DocMind AI (custom) |
|---|---|---|---|
| Precision | Variable, fatigue-dependent | ~85-92% | 95% baseline / 99.5% with fallback |
| Cost per document | $50 to $1,000+ analyst time | $2 to $10 / doc, plus QA | Under $0.30 per document at scale |
| Processing time | 4 to 8 hours per document | Minutes per page, but heavy QA | ~4.22 minutes per document end-to-end |
| Custom parameters & schemas | Possible but inconsistent | Limited, generic templates | 100+ parameters, fully client-defined |
| Table & chart extraction | Time-consuming, error-prone | Often loses context | Native handling with relationship preservation |
| Validation & cross-reference | Manual second-reviewer | Not provided | Multi-layer business rules + derived values |
| Scalability | Linear with headcount | Capped by template fit | Unlimited compute scaling, dollar-priced |
| FDTA 2027 readiness | Manual restructuring needed | Retrofit required | Structured manifest by design |
| Audit trail | Inconsistent, paper-based | Limited logs | Full prompt + source + override log |
| Integration depth | Manual data entry into systems | Standard REST API | REST API, direct DB write, EMMA/MAC/MSRB feeds |
| Compliance posture | Process-dependent | Generic, not regulated-specific | ISO 9001, ISO 27001, HIPAA per project, MSRB-aware |
Two organization-wide certifications are active and audited. Other compliance regimes apply per project, scoped to the client's regulatory environment. We list both categories here so prospective clients know what is automatic and what needs scoping.
Document AI is not the right answer for every use case. We turn down projects that do not fit. Below is the honest picture of where DocMind AI is the right partner and where you should look elsewhere.
You process hundreds to thousands of long PDFs per month - bond official statements, regulatory filings, financial reports, scanned forms - and manual review is the bottleneck.
You operate in a regulated environment (US municipal finance, banking, insurance, healthcare, government) and need an audit-ready system, not a chat-with-PDFs prototype.
You are a US municipal advisor or underwriter facing the FDTA 2027 machine-readable mandate and need structured output by design, not a 2026 retrofit.
You need custom extraction schemas with 50+ parameters per document, validation rules, and cross-reference checks - not generic OCR templates.
You want a partner that builds the pipeline with you, ships it to production, then operates and tunes it - not a self-serve SaaS or a deck-and-handoff consultancy.
You need to integrate with EMMA, MAC, MSRB, or similar issuer / regulator data sources, plus your internal CRM, GRC, or treasury systems.
You need a shrink-wrapped SaaS for under $200 per month. We build production custom pipelines, not self-service tools.
You want a chat-over-PDFs prototype with no production path. Generic LLM chat tools cover that for a fraction of the cost.
Your document volume is under 50 documents per month with simple extraction needs. Generic OCR or manual review is more cost-effective at that scale.
You want to fully automate decisions in regulated environments with no human in the loop. MSRB G-42 and similar rules require analyst sign-off, and we will not deliver against that constraint.
You need real-time order routing, payment processing, or e-commerce orchestration. Those are different categories of system - we build them on other engagements, not as DocMind AI.
You expect 100% accuracy with no fallback. We design for 95% baseline, 99.5% with fallback, and explicit human review for edge cases. Anyone promising 100% is selling you a problem.
A pre-formatted summary for fast reference and citation. Each card is a self-contained answer to a frequent question.
DocMind AI is a production AI document processing and data intelligence capability built by Kyotu Technology. It covers two parallel tracks: extraction of structured data from long PDFs (bond official statements, regulatory filings, scanned forms) and generation of new documents (POS, FOS) from source data, historical templates, and rule packs. Live in production at the Municipal Advisory Council of Texas and SAMCO Capital Markets.
Municipal bond Preliminary Official Statement (POS), Final Official Statement (FOS), Notice of Sale, Bond Resolution / Order, Remarketing Memorandum, plus continuing-disclosure filings, financial reports, dental insurance plans, payor schemas, and adjacent regulated documents. Custom schemas per client - we treat document type as a configuration, not a hardcoded module.
Production reference at MAC of Texas: 5,236+ documents parsed, 91 documents per day continuous throughput, 4.22 minutes per document average end-to-end. Architecture is dollar-priced compute - throughput scales linearly with budget, not headcount. Multi-million-document archive deployments are pattern-fit, scoped per project.
$0.53 per document at MAC of Texas in production. Below $0.30 per document achievable for higher-volume deployments with optimized prompts. Compared to manual analyst review at $50 to $1,000+ per document and generic OCR plus QA at $2 to $10 per document. Cost is dollar-deterministic - we publish actuals, not estimates.
Production deployments in Austin, Texas, US: Municipal Advisory Council of Texas (the official State Information Depository for Texas since 1995) and SAMCO Capital Markets (a 100%-employee-owned boutique investment bank founded 1987). Plus a 4+ year US healthcare-tech engagement (anonymized) for dental insurance verification across multiple payor portals.
Kyotu Technology is ISO 9001 and ISO 27001 certified organization-wide. DocMind output is structured for FDTA 2027 by design. Workflows are aware of SEC Rule 15c2-12, MSRB G-17 and G-42, HIPAA per project. Audit trail covers every prompt, source citation, validation rule, and analyst override. DUNS 679252803.
Twelve questions we get asked most often by municipal advisors, broker-dealers, banks, healthcare-tech companies, and enterprise document teams scoping a DocMind AI deployment.
Off-the-shelf tools (V7 Labs AI Agent for Public Finance, DBC Data Agent, MuniBonds.ai) cover generic municipal-finance extraction with templates designed for broad market fit.
DocMind AI is built around client-specific schemas, custom validation rules, deep integration with EMMA, MAC of Texas, State Comptroller, and similar issuer data, plus the parallel generation track. We compete where domain depth and custom integration matter more than self-serve onboarding. We do not compete on per-seat SaaS pricing.
Yes. The ingestion layer handles text-based PDFs, scanned images, and mixed-format collections. OCR preprocessing, layout normalization, and table recognition are part of the standard pipeline. Bond official statements often combine native text and scanned exhibits in the same document - DocMind AI handles both in a single pass.
Baseline precision is 95% on first iteration. With confidence-thresholded fallback validation - re-extracting low-confidence fields with alternative prompts and cross-reference checks - we routinely reach 99.5% in production on bond document parameters.
The 0.5% remainder is routed to a missing-fields panel for analyst review, not silently failed. We design every deployment around this loop, not around an unrealistic 100% promise.
Document assessment in 48 hours. Pilot build over 2 to 4 weeks. Validation and tuning in-pilot. Full production deployment in 4 to 8 additional weeks. Total: 6 to 12 weeks for most municipal-finance deployments.
Healthcare-tech and high-volume operational deployments scope per project depending on integration complexity and compliance posture.
Kyotu Technology is ISO 9001 and ISO 27001 certified organization-wide. HIPAA, SOC 2 Type II, and GDPR / EU data residency are scoped per project to fit the client's regulatory environment. Our 4+ year US healthcare-tech engagement runs HIPAA-aware end-to-end. We will not claim certifications we do not hold - on-paper compliance is reviewed and signed before any client engagement begins.
Two components: a fixed-price pilot build (typically $25k to $80k for the 2 to 4 week pilot) plus ongoing operations costs tied to document volume.
Operations are dollar-deterministic: at MAC of Texas the unit cost is $0.53 per document, and below $0.30 per document is achievable for higher-volume deployments. Some clients prefer a managed-service monthly fee covering volume + maintenance - we structure either model.
Yes. EMMA, MAC of Texas, and Texas State Comptroller are already integrated in our reference deployments. For internal systems, we deliver REST API, direct database write (PostgreSQL, MongoDB), structured-file export, or message-queue handoff. CRM, GRC, treasury, and practice-management integrations are scoped per project. We do not gate integrations behind a marketplace.
Documents are processed within client-controlled infrastructure or through Kyotu-operated tenants under DPA / BAA contracts as required. We use commercial LLM endpoints with no-training data terms (OpenAI Enterprise, Azure OpenAI). Client documents are never used to train any model. For regulated environments where this is not acceptable, we design air-gapped or on-prem deployments with self-hosted open-weight models.
Yes, per project. Our default is managed cloud (Azure or AWS in the client's preferred region). For regulated environments that require it, we deploy into the client's private cloud, virtual private cloud, or on-prem with self-hosted models. EU data residency for GDPR, US data residency for sensitive defense or healthcare contexts - both are deliverable, scoped during pilot.
The Financial Data Transparency Act requires US municipal issuers, broker-dealers, and underwriters to file continuing disclosures in machine-readable format by 2027. The federal mandate covers structured tagging, schema-aligned data, and validated submission to repositories.
DocMind AI emits structured-data manifests as a primary artifact alongside every generated or extracted document. Clients deploying DocMind today are FDTA-ready by design, not retrofit-ready.
Both. The Extract track pulls structured data out of existing PDFs (live at MAC of Texas). The Generate track composes new documents - Preliminary and Final Official Statements - from source data, historical templates, and rule packs (under active development at SAMCO Capital Markets).
Same architecture, two directions. Many clients deploy both for full document-lifecycle automation: extract data from past issuances to inform the next one, generate the new POS / FOS, and validate against MSRB / SEC rules in a single pipeline.
Two layers of defense. First, confidence-thresholded fallback automatically re-extracts low-confidence fields with alternative prompts and cross-reference rules - this is the path that takes 95% baseline up to 99.5% in production. Second, fields that remain below threshold are routed to a missing-fields panel for analyst review with the original document context, the extracted value, and the validation rule that flagged it.
Every override is logged to the audit trail. We do not silently fail or hallucinate values that look plausible. Where MSRB G-42 requires the analyst to personally verify, the workflow enforces that step.