Here's the Exact Architecture That Processes 847 Invoices in Seconds
Multi-Agent Document Processing Pipeline for High-Volume Financial Intelligence
---
A CFO asks: "Which suppliers raised prices more than 10% this year?"
Answering that question properly requires checking 847 invoices. Old approach: export to Excel, build a pivot table, cross-reference manually — three hours minimum, an analyst's Friday afternoon. With Stralevo, the same question returns a sourced answer in seconds for a standard indexed query.
What follows is an explanation of exactly what happens in those seconds. Not in engineering terms — in the terms a CFO evaluating financial AI needs to understand: what each component does, why it matters for accuracy, and what it means that every answer arrives with a citation to the exact document and line item that produced it.
---
Most Financial AI Has a Fundamental Design Problem
Most "AI for finance" products are a chat interface placed on top of an existing accounting system. The query runs through a language model that has access to the same 3–5 fields your accounting software captured from each invoice: the date, the amount, the VAT number, the vendor name, the description. The answer arrives quickly because there is so little data to search.
Speed is not the problem. The issue is that the 35 fields your accounting software ignored — serial numbers, warranty terms, delivery references, payment conditions, contracted price schedules, unit-level pricing — are not in the index. When the language model answers "which suppliers raised prices more than 10%?", it is answering from 10–30% of the data available in your own documents. And in most systems, nothing verifies whether the answer it generated is actually correct before it reaches you.
Stralevo's architecture addresses both problems: full data capture before any query runs, and independent answer verification before any answer is delivered.
---
What Actually Happens in Those Seconds
Reading the Documents
Before any query runs, Stralevo's SightCapture technology reads every document completely — not just converting images to text (which is what standard software does), but understanding the document as a human reader would. It captures every field in every document: serial numbers embedded in mixed-content fields, warranty terms in small print, unit prices in irregularly formatted tables, handwritten annotations on scanned receipts.
Standard accounting software reads 3–5 fields per invoice to satisfy reporting requirements. SightCapture reads all 15–40. The additional 35 fields — the ones that answer questions about warranties, contracted pricing, delivery terms, and supplier relationships — are captured the first time a document is processed and indexed for every query that follows. That is why the query across 847 invoices is fast: the documents do not need to be read again at query time. They are already completely indexed.
SightCapture works with 50+ document formats: PDFs, scanned images, electronic invoices, contracts, bank statements, receipts. Every format, every field.
Understanding the Question
ContextUX, the input layer, processes what the CFO actually asked — in natural language, by text, voice, or video — and translates it into a structured query. "Which suppliers raised prices more than 10%?" becomes a query against the unit price fields across 847 invoices, cross-referenced against contracted price schedules, filtered for increases above the threshold.
Role adaptation works the same way. A bookkeeper asking about invoice processing sees bookkeeping workflows. A CFO asking about quarterly performance sees executive summaries. An auditor checking compliance sees audit trails. Same system, same data — the relevant view arrives automatically based on who is asking.
Searching Across Every Document
Needle Finder, the cross-document query engine, runs the search. A single question can span hundreds of documents simultaneously: 234 invoices cross-referenced with 47 supplier contracts to identify cases where the invoice unit price exceeds the contracted rate. Needle Finder resolves those relationships across the complete document set and returns the relevant results.
Response times for queries like this depend on data volume, query complexity, and system load. For standard queries against indexed data — like the supplier pricing question — the result arrives in seconds. For complex cross-document analysis spanning thousands of unindexed documents, the result may take minutes. Stralevo is specific about this because the alternative — claiming uniform speed regardless of query type — is a claim that does not survive the first real-world test.
Reconciling the Transactions
Smatched, Stralevo's reconciliation engine, handles the transaction matching problem that breaks most automated financial systems. When one payment from a buyer covers multiple invoices. When multiple partial payments cover a single invoice. When payment references do not match invoice numbers exactly. When a vendor issues a credit note that offsets part of an outstanding invoice.
Bank reconciliation still requires a skilled finance professional in most organizations from this N:M (many-to-many) matching problem — one payment against many invoices, many payments against one invoice, and every combination in between. Smatched resolves it automatically, flagging genuine exceptions rather than treating every non-standard transaction as one. It took 18 months of R&D to build, and is patent pending — because the matching problem it solves is genuinely hard. The result is that the reconciliation work currently occupying hours of your team's time each month happens automatically, with exceptions surfaced for human review.
Verifying Before Delivery
Inside the pipeline, the Verification Layer is the most important component — and the least visible to the person asking the question.
In most AI systems, the model that generates the answer and the model that checks the answer are the same — or there is no check at all. Stralevo's architecture separates the two processes entirely. After the Reasoning Engine generates an answer, an independent Verification Layer checks it. A separate agent, which does not know what the first agent concluded, independently confirms that the answer matches the source data.
"Zero-hallucination tolerance" means in practice: not that the AI is incapable of generating a wrong answer, but that the architecture catches wrong answers before they reach the CFO. Every answer that passes the verification check arrives with a citation to the exact source: which document, which page, which line item. If the answer cannot be verified against a source, Stralevo says so rather than delivering an unverified conclusion.
Delivering with Sources
VibeFlow, the output layer, delivers the answer in the format most useful to the person who asked — a structured summary for the CFO, a detailed breakdown for the analyst, an audit trail for the compliance team. Every answer includes the source citations that prove it.
An answer like "which suppliers raised prices more than 10%?" arrives with the invoice numbers, the line items, and the contracted rates that produced each figure. A CFO presenting that analysis to a board does not need to say "the AI said so." They can say "invoice number 847 from Supplier X, dated March 12, shows a 14.3% unit price increase against the contracted rate in the supply agreement signed January 2024."
That is a fundamentally different kind of financial intelligence than an unsourced AI summary.
---
Why Source Citations Are Not Optional in Finance
JPMorgan's COIN system reduced 360,000 hours of annual contract review to seconds by reading documents directly rather than relying on what humans had previously summarized from them (Bloomberg, 2017). The lesson from that deployment: document-native intelligence — where the system reads the original source rather than working from a secondary extract — is both faster and more complete than alternatives.
Samsung engineers pasted semiconductor source code into ChatGPT three times in a single month because no sovereign, document-native alternative existed that could answer their questions from their own documents. The combination of speed, completeness, and sovereignty is what that incident made visible as a gap.
EU AI Act Article 13 formalizes the requirement for financial AI: high-risk AI systems used for business decisions must provide transparency sufficient for users to understand and verify the output. Financial AI that drives purchasing decisions, supplier management, tax filings, and board reporting is classified as high-risk. Systems without source citations will face a compliance retrofit when enforcement begins in 2026. The architecture that builds in source citations from the start is compliant by design.
Regulators are now formalizing the same standard any experienced auditor would apply: show your working. The accounting profession has operated on that principle for decades. The AI systems used in finance should meet the same standard.
---
The Trust That Compounds With Accuracy
When answers arrive with source citations, trust in the system builds at a different rate than trust in systems without verification.
Any CFO who verifies three answers against their source documents and finds them accurate will trust the fourth answer without checking. That earned trust changes how financial intelligence is used: not as a starting point for manual verification, but as a final answer that drives decisions. The time saved is not just the query time — it is the verification time that was previously unavoidable for every AI-assisted figure used in a financial report.
Its opposite is equally real. A system without source citations accumulates unearned trust — the pattern of unchallenged answers that eventually produces an undetected error in a decision that cannot be corrected. The architecture that prevents that pattern is the one that requires every answer to be proven before delivery.
---
What Becomes Possible
Once all 40 fields of every financial document are indexed, verified, and queryable, the questions that currently require an analyst's afternoon become immediate. Supplier price increases across 847 invoices: seconds. Warranty coverage on equipment purchased in the last three years: seconds. All invoices from a supplier where the unit price exceeded the contracted rate: seconds. The contracts due for renewal in the next 90 days: surfaced automatically before the CFO asks.
Proactive intelligence — Stralevo identifying anomalies, duplicate invoices, overcharges, and renewal deadlines before a question is asked — is only possible because every document is fully indexed at ingestion. A system that retrieves data on demand can only answer questions that are asked. A system with a complete, current document index can identify the questions that should have been asked but have not been yet.
The last question for any CFO evaluating financial AI is this: when was the last time an AI-generated financial figure you acted on included a citation to the exact source document, page, and line item? If the answer is "never," that is the gap this architecture was built to close — on your infrastructure, sovereign, available today on any accounting software you already run.