Method and system for enhanced LLM analysis with application to accounting data

Abstract

A method and system for enhanced financial data processing using large language models (LLMs) to transform PDF statements into structured accounting data. The invention preprocesses statements with line numbering, character position tagging, and data partitioning to overcome LLM context limitations. The system extracts transactions, reconciles them with statement balances, normalizes vendor information, and generates enhanced output files compatible with standard accounting platforms. This enhanced transaction data collapses reconciliation and categorization times in downstream accounting systems. The system creates a robust vendor-transaction database supporting direct analysis or simultaneous application of multiple accounting models.

Description

APPLICATION FOR UNITED STATES LETTERS PATENTMETHOD AND SYSTEM FOR ENHANCEDLLM ANALYSIS WITH APPLICATION TOACCOUNTING DATAInventors: William Lester DyeDanamichele Brennen DyeDye Research Inc7461 Whileaway RdPark City, UT, 84098435637-6721METHOD AND SYSTEM FOR ENHANCED LLMANALYSIS WITH APPLICATION TO ACCOUNTINGDATA CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of U.S. Provisional PatentApplicationNo. 63 / 645011,filed May 9, 2024, entitled ”Method and System for Enhanced LLM Analysis with Appli-cation to Accounting Data,” which is incorporated herein by reference in its entirety.FIELD OF THE INVENTION

[0002] This invention pertains to the fields of data processing and data analysis, specificallythe preprocessing of semi-structured financial data to enhance the analytical capabilities oflarge languagemodels (LLMs). In particular, it addresses the previously intractable prob-lem of accurately extracting accounting data from diverse and heterogeneous PDF bank,credit card, and other statements and creating a coherent vendor-transaction database thatserves as an objective historical record. This new enhanced transaction data is used di-rectly for analysis or for import into all accounting systems fully replacing bank data. Thisresults in remarkable encoding, categorization and reconciliation time reductions and ac-curacy improvements. The invention reconciles extracted data to known beginning andending balances, thereby ensuring both accuracy and completeness. It allows for mul-tiple subjective accounting models to be applied to the same objective transaction data,overcoming limitations of traditional accounting systems that apply only one model to acompany or individual. Unlike prior approaches that require partial human intervention,this invention enables a fully automated pipeline.BACKGROUND OF THE INVENTION

[0003] Portable Document Format (PDF), standardized as ISO 32000, is a file format de-veloped in 1992 by Adobe to present documents, including images and text formatting,regardless of the operating system, hardware, or application software being used. ThePDF is accepted globally, and by the United States Internal Revenue Service, as the state-ment of record for a financial account.

[0004] PDF has become the global standard for financial statements for several critical rea-sons: all banks provide PDFs while only a small percentage support direct connections;PDFs offer long-term availability, superior security, and complete data; and PDFs serve aslegally valid source documentation required for audit trails and financial recordkeeping.In contrast, alternative formats like OFX (Open Financial Exchange) have significant limi-tations, with only approximately 5% of banks worldwide and 15% of US banks supportingthe format. Additionally, OFX connections are typically limited to just 90 days of history,pose security risks through third-party connections, and do not map to specific monthlystatements, causing reconciliation issues.

[0005] Traditional methods of extracting and processing financial data from these PDFstatements typically involve manual intervention, resulting in inefficiencies, inaccuracies,and incomplete data reconciliation. Current systems strugglewith the vast variety of state-ment formats and layouts,making the extractionprocess error-prone and time-consuming.

[0006] Furthermore, existing technologies face challenges in reconciling the extracted datawith beginning and ending balances, ensuring data completeness, and standardizing ven-dor information across different statement formats. These limitations hamper the effec-tiveness and reliability of financial data analysis and accounting processes, contributingto the decline in US small business use of accounting software from 85% to 68% in recentyears.SUMMARY OF THE INVENTION

[0007] This invention is a computer utility that uses artificial intelligence (AI) to replaceand improve traditional, manual bookkeeping activities. The PDF is accepted globally,and by theUnited States Internal Revenue Service, as the statement of record for a financialaccount. This invention utilizes modern AI LLMs (large language models) to read thesePDF files, convert them into transaction-level data that is enhanced with accurate vendorinformation and organized into pre-reconciled statement groups. This process is faster,more accurate, and less expensive than traditional manual and electronic methods. Thisenhanced data can be used to do analysis directly or imported via industry standard OFX,QBO or CSV formats into any accounting system. The statement-reconciled enhanceddata yield remarkable savings in time in categorization and reconciliation in downstreamaccounting systems by arming their rule based systems with accurate and complete data.Furthermore, this new approach makes it possible to directly create a coherentvendor-transaction database that serves as an objective historical record. This data foun-dation can be used to create financial statements such as income statements and balancesheets for specific accountingmodels using subsequentAI pipelines. Previous approachesrequired manual intervention to identify vendors and reconcile transactions prior to directinsertion into one unique accounting model. This invention allows for multiple subjectiveaccounting models or simple reports to be applied to the same objective transaction data.The invention allows mathematically correct transaction transformations such asredaction, aggregation and renaming that reduce downstream analysis cost or improvesecurity and privacy without loss of reconciliation.

[0008] This computer software utility invention includes the following main steps:1. The pre-processing stage where the software converts the PDF into semi-structuredtext. 2. Structured injection of line numbers and tagging of numerical data and headerswithcharacter offset positions.3. Partitioning the data into sections that can be processed concurrently by the LLM.4. Assemblage of the processed data into a coherent whole.5. An advanced interface to the LLM that interprets the text and returns key compo-nents of the statements and all transaction details.6. The reconciliation of those transactions to a beginning and ending balance, and thefixing of any errors that may arise using the LLM.7. Analyzing the memo fields of each transaction and larger context using the LLM andconverting them to a rationalized, standardized format for the vendor / payee / nameand / or the intra-account transfers.8. Further optional input from the user in the form of ”Directive” files to transform thetransaction data such as renaming payees and aggregating transactions from specificvendors. 9. Output in the form of common financial files and documents such as OFX 2.0 com-pliant accounting system upload files, QBO format files, and CSV formats.10. Optionally applying an accounting model to bypass the accounting system to di-rectly generate financial statements such as income statements and balance sheets.

[0009] This process can be thought of as ’Preconditioning, Processing, Injection, and Trans-formation’ of the data. Preconditioning allows advanced partitioned Processing to opti-mize LLM context usage. Vendor and other information are extracted (not available inOFX and other electronic bank data) and this is Injected into standard data formats. Op-tionally these enhanced data transactions are Transformed into transaction sets that retaincoherence to account balance. The system is designed to be modular and extensible, al-lowing for future enhancements and additional features as needed. These elements arenovel and non-obvious.BRIEF THE

[0010] how theuser outputin

[0011] showinghow data

[0012] enhances

[0013] andrules method.

[0014] demon-stratingDISTRIBUTED COMPUTING DATA FLOW DESCRIPTIONFigure 1: Data Flow Diagram of the Utility Application

[0015] Figure 1 illustrates how the user computer sends PDFs and directives to the util-ity application, which produces output in the form of enhanced accounting system inputfiles (QBO, OFX, CSV). The utility application interacts with file handling systems, andapplication processing servers to process the data and produce the desired output.

[0016] The system includes:• User interface components (U 000) for file uploads and directive inputs• Processing nodes (BC 001, BS 001, TC 000) that handle various aspects of the dataprocessing •Load balancers and job queues that manage the distributed workload• Storage systems (FileStack FS) for both input and processed data• External AI systems that assist in the data analysis

[0017] Each component in the systemperforms specific functions in a sequence that ensuresefficient processing of the incoming data and generation of accurate output files.PRECONDITIONING OF DATA FOR LARGE LANGUAGE MODELANALYSIS Figure 2: Post Injected Statement Data in text block format

[0074]

[0075] Card Trans Post Reference Number Description Credits

[0131] Charges

[0150]

[0076] Ending Date Date

[0077] 1530 08 / 08 08 / 08 74465396X7EH872XQ CASH BACK REDEM 100.00

[0133]

[0078] 1530 08 / 09 08 / 09 24492166Y000JQN4G GENESYS 329.00

[0150]

[0018] Figure 2 illustrates the post-injected statement data. The data is preconditioned toinclude injected line numbers and optionally numerical data elements tagged with char-acter offset positions.

[0019] Raw transaction data from financial institutions often appears in inconsistent for-mats across different statements, making rule creation nearly impossible in traditionalaccounting systems. Consider these examples of description information for the samevendor appearing in different formats:1. PIN THE HOME DEPOT 1126 GREER2. Debit Card Purchase 011421 THE HOME3. HOMEDEPOT.COM 800-430-3376 GA

[0020] This invention’s AI processing standardizes all these variations to a single normal-ized vendor name: ”The Home Depot”. This normalization is performed automaticallywithout the need for complex manual rule creation, which dramatically improves effi-ciency and reduces the potential for errors incategorization.Figure 3: Line Numbering Transformation Example

[0001] Redacted Name

[0002] Redacted Address

[0003] REDACTED CITY STATE, ZIP

[0004]

[0005]

[0006]

[0007]

[0008] BUSINESS ACCOUNT NUMBER: *** 1723 Statement Range:

[0009]

[0010] Account Summary August 1, 2024 --- August 31, 2024 (31 days)

[0011]

[0012]

[0013]

[0014] Beginning Balance on Aug 1, 2024 $50.00

[0015]

[0016] Funds Added (Deposits and Credits) +$1,458.85

[0017]

[0018] Funds Moved (Withdrawals, Fees, and Other Debits) -$1,508.85

[0019]

[0020]

[0021] Ending Balance on Aug 31, 2024 $0.00

[0022] ...

[0050] Posted Date Description Category Amount Balance

[0051]

[0052] 8 / 3 / 2024 CASH APP*REDACTED, San Francisco, CA Debit -$50.00 $0.00

[0053]

[0054] 8 / 4 / 2024 SQ *ID - GYKOMRIW, SAN FRANCISCO, CA Credit +$60.27 $60.27...

[0021] The line numbering transformation enhances the data for LLM analysis as demon-strated in Figure 3. This structured approach to data preparation enables more accurateand efficient analysis by the LLM, allowing for precise reference to specific locationswithinthe statement and facilitating sectional processing of large documents.ADVANTAGES

[0022] The invention provides significant advantages over traditional methods of process-ing financial data as demonstrated by comparative analysis in the tables and figures pre-sented in the Quantitative Performance Analysis section.

[0023] Specifically, the advantages include:1. Improved accuracy in data extraction, decreasing reconciliation error rate from 1 in5 statements to 1 in 100 statements.2. Indexing of lines allows for advanced partitioning of data into sections that can beprocessed concurrently by the LLM, enhancing processing speed and efficiency.3. Indexing of lines allows data to be partitioned to fit specific LLM context windowsto reduce cost or improve performance.4. Indexing of lines allows for weaker less expensive LLMs to be used, to obtain com-plete or partial information cheaply. If not complete, then a sequential process of in-creasingly powerful LLM requests can utilize previous results to provide enhancedLLM cost performance.5. Allows processing of otherwise intractable statements with thousands of transac-tions through efficient partitioning and recombination techniques.6. Enhanced human time processing speed, with up to 95% reduction in reconciliationtime and 60% reduction in categorization time compared to traditional methods.Enhancedhuman timeprocessing speed, with ability to encode>50,000 transactionsper hour compared to manual encoding at approximately 30 per hour with humanencoding of statements.Simplified downstream accounting system rule management, maintaining a perfect1:1 relationship between vendors and rules, unlike traditional methods that requireincreasingly complex rule systems.Enhanced data integrity through unique identifiers for each transaction, preventingduplication or transaction ’lock-out’ in accounting systems.Transformation capabilities of vendor remapping and aggregation, which may col-lapse transaction count by grouping or aggregating high frequency transactionswhoseindividual elements are not useful for downstream accounting systems. Repeated’Starbucks’ transactions may be aggregated into a single monthly transaction.Enhanced scalability through partitioning of data into sections that can be processedconcurrently by the LLM, making large statements feasible for analysis. A typicalLLM such as OpenAI GPT 4 may have a 120,000 token context window but a small4096 token limit for output. Partitioning allows extraction of transaction groups thatdo not exceed output token limits or that optimize LLM performance or cost.Elimination of manual intervention in the reconciliationprocess, reducing labor costsand potential for human error.Standardization of payee information across different statement formats, improvingdata consistency and analysis capabilities.Generalization of accounting data input is now possible, allowing for the use of anyPDF statement as input. If a statement can be transformed into a synthetic statementwith beginning and ending balances and transactions then this invention can processit. The invention has processed Amazon, Costco, PayPal and many other sources oftransaction data. This is a significant improvement over traditional methods thatrequire large labor investment. In the case of Amazon, humans have difficulty iden-tifying product or sub-vendor names. This invention can identify the product orsub-vendor names consistently and accurately.15. The systemworkswithoutmodification for language andglobal currency bank state-ments, allowing for international use.

[0024] The significant cost advantages of this invention are substantial when viewed in anational context. Bookkeepers and accountants charge an average of $2.00 per transactionnationally1. With this AI-based invention, assuming time is valued at $100 / hr and pro-cessing 50 transactions per statement, labor cost drops to approximately $0.16 per trans-action, plus only $0.03 per transaction for the service fee ($1.50 per statement). A $0.19per transaction total cost gives a significant cost savings.:• Before: $2.00 / transaction• After: $0.19 / transaction– Human Cost: $0.16 / transaction– AI Cost: $0.03 / transactionQUANTITATIVE PERFORMANCE ANALYSIS

[0025] The studies presented provide a comprehensive quantitative analysis of the inven-tion’s performance advantages across different accounting systems and account types.

[0026] Manual data entry involves manually entering transaction data into a CSV, whichis time-consuming and error-prone. This CSV is then imported into QuickBooks / Xero1Asample of per transaction rates at 24 bookkeepingfirms across theUSgives amean of $2.38 / transactionand a median of $1.96 / transaction.identically to the Electronic approach. One could of course enter directly into the registerdirectly but this is generally slower as rules are skipped requiring the user tomanually cat-egorize each transaction. Most banks and credit unions globally do not provide electronicversions of their statements.

[0027] The electronic approach uses bank feeds to import transaction data directly intoQuickBooks / Xero. This method avoids the manual entry of name, description, amount,and date information but requires vendor identification, categorization, and reconciliationidentical to the manual approach. QuickBooks / Xero provides rule-based mechanisms toassist in this process, but they are slow because of the nature of the name and descriptiondata.

[0028] Finally, theAI-enhanced pre-processing approach standardizes and reconciles trans-action data before it enters QuickBooks, enabling perfect rule application and trivial rec-onciliation. In these case studies we time the complete processing of a statement from PDFor electronic data to accounting system reconciliation.

[0029] The AI-based processing methods dramatically outperform both manual entry andelectronic import methods consistently across both checking and credit card accounts inthe popular accounting systems Xero and QuickBooks.

[0030] For checking accounts, the AI Preconditioned method reduces processing time by83.9% compared to manual entry in Xero (from 84.36 minutes to 13.57 minutes) and by87.2% in QuickBooks (from 80.31 minutes to 10.30 minutes). The Raw AI Component,which represents the invention processing time is just 0.35 minutes. This demonstratesthe efficiency of the AI technology itself, while the AI Preconditioned method is the AIprocessing time plus the greatly reduced human time.

[0031] For credit card accounts, which typically involve more transactions per statement,the performance improvements are equally impressive. The AI Preconditioned methodreduces processing time by 86.2% compared to manual entry in Xero (from 208.64minutesto 28.73 minutes) and by 89.3% in QuickBooks (from 216.55 minutes to 23.17 minutes).The Raw AI Component time is just 1.50 minutes, highlighting the dramatic reduction inmachine processing time achieved by our technology.

[0032] The Raw AI Component represents only the machine processing time for the ini-tial PDF analysis, excluding subsequent human-dependent steps like categorization in ac-counting timesand software.

[0033] average cost perpro-cessing. savings foreliminating hu-man dramatically re-duces theFigure 4: Cumulative growth comparison showing new rules created for each new vendoradded using electronic and AI-based processing methods. The dashed line represents aperfect 1:1 relationship.Rule Growth Comparison40 s le30 uRweN evita20 lumuC 10 0 010 20 30 40Cumulative New VendorsElectronic AI Perfect

[0034] Figure 4 illustrates another key advantage of the AI-based approach: the main-tenance of a perfect 1:1 relationship between vendors and rules. Traditional electronicmethods require an increasing number of rules as more vendors are added, leading toexponential growth in complexity. With just 22 vendors, the electronic method requires43 rules, while the AI method maintains exactly 22 rules. This simplification is critical forscalability and long-term maintenance of the system.

[0035] The detailed time breakdowns in the tables reveal that the most significant timesavings occur in the CSV conversion phase (tcsv) and reconciliation phase (trec). For ex-ample, in credit card processing with QuickBooks, reconciliation time drops from 25.57minutes with manual and electronic methods to just 0.52 minutes with AI Preconditionedprocessing, a 98% reduction.

[0036] It is important to note that while our AI approach significantly reduces the timerequired for categorization and reconciliation, it does not completely eliminate the needfor human review within the double-entry accounting system. The AI Preconditionedmethod represents the complete process including the reduced but still necessary humansteps for final verification and accounting system integration. Even with these remaininghuman steps, the overall time and cost savings are substantial.Table 1: Processing Details for Checking Accounts in XeroMethod Time Cost Cost / TransManual Entry 84.36140.58 2.38Electronic Import 31.85 2.25 20.97 6.38 53.08 0.90AI Preconditioned* 13.57 0.35 10.53 0.43 27.11 0.46Raw AI Component† 0.35 0.35 0.00 0.00 5.08 0.09* AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time without accounting for subse-quent human stepsNote: All times in minutes. 59 transactions over 3 statements. Costs based on $100 / hr laborrate. RECONCILED CSV FILES

[0037] Derivatives of the OFX Standard are the common format for files that upload fi-nancial data into accounting systems. Intuit’s QuickBooks accounting package calls itsstandard upload file a ”QBO file” format with a ”.qbo” as the end of the file name; whenone opens a file with a ”.qbo” at the end of the file name, the QuickBooks application auto-matically opens up and begins importing the data in that file. Almost all other accountingsystems on the market today will upload data using the OFX file standards and / or in aTable 2: Processing Details for Checking Accounts in QuickBooksMethod Time tcsv tcat trec Cost Cost / TransManual Entry 80.31 54.75 17.58 5.72 133.84 2.27Electronic Import 27.80 2.25 17.58 5.72 46.34 0.79AI Preconditioned* 10.30 0.35 6.68 1.02 21.66 0.37Raw AI Component† 0.35 0.35 0.00 0.00 5.08 0.09* AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time without accounting for subse-quent human stepsNote: All times in minutes. 59 transactions over 3 statements. Costs based on $100 / hr laborrate. Table 3: Processing Details for Credit Card Accounts in XeroMethod Time tcsv tcat trec Cost Cost / TransManual Entry 208.64 148.98 44.45 12.20 347.72 1.39Electronic Import 62.65 3.00 44.45 12.20 104.41 0.42AI Preconditioned* 28.73 0.75 24.72 0.27 53.88 0.22Raw AI Component† 1.50 0.75 0.00 0.00 7.25 0.03* AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time without accounting for subse-quent human stepsNote: All times in minutes. 250 transactions over 4 statements. Costs based on $100 / hr laborrate. Table 4: Processing Details for Credit Card Accounts in QuickBooksMethod Time tcsv tcat trec Cost Cost / TransManual Entry 216.55 148.98 39.00 25.57 360.92 1.44Electronic Import 70.57 3.00 39.00 25.57 117.61 0.47AI Preconditioned* 23.17 0.75 18.90 0.52 44.61 0.18Raw AI Component† 1.50 0.75 0.00 0.00 7.25 0.03* AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time without accounting for subse-quent human stepsNote: All times in minutes. 250 transactions over 4 statements. Costs based on $100 / hr laborrate.Table 5: Comparative Analysis of Processing Methods Across All Account TypesChecking Credit CardMethod Xero QuickBooks Xero QuickBooks Avg. Cost / TransManual Entry 84.36 80.31 208.64 216.55 1.59Electronic Import 31.85 27.80 62.65 70.57 0.52AI Preconditioned* 13.57 10.30 28.73 23.17 0.24Raw AI 0.35 0.35 1.50 1.50 0.04Component†*AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time without accounting for subse-quent human stepsNote: Values represent processing time in minutes. AI Preconditioned includes human timerequired for a complete accounting workflow. Costs based on $100 / hr labor rate.Table 6: Time Efficiency Ratios Compared to AI-Based ProcessingChecking Credit CardMethod Xero QuickBooks Xero QuickBooksManual Entry 241.0 229.5 139.1 144.4Electronic Import 91.0 79.4 41.8 47.0AI Preconditioned* 38.8 29.4 19.2 15.4Raw AI Component† 1.0 1.0 1.0 1.0* AI Preconditioned represents the complete process including necessary human steps in ac-counting systems† Raw AI Component shows only the machine processing time, used as the baseline for compar-ison Note: Values represent howmany times slower eachmethod is compared toRawAIComponentprocessing.”.csv” (Comma-separated values) upload file format. This invention improves the infor-mation content of these input files (QBO, OFX and CSV files) for the accounting packageuser, with vendor names, transaction analysis, check numbers and reconciliation factors.Reconciliation factors are those things that help the user to reconcile, or balance, a state-ment, or group of statements, to a beginning balance and ending balance; i.e. if a statementis reconciled, then every transaction has been accounted for in the additions and subtrac-tions that occur within a statement, or group of statements.The computer utility creates output in the form of reconciled QBO, OFX and CSVfiles which are by definition acceptable to all accounting packages. This eliminates anypossible issues with downstream changes in accounting systems such as QuickBooks,Xero, Freshbooks, etc.10.1 ADVANTAGES

[0038] The transformations and enhancements made by this invention solve known prob-lems with traditional QBO, OFX and CSV files and equivalent electronic feeds that arecreated by banks and credit card companies and these have been persistent for decades,much to the end users frustration.

[0039] This computer software utility creates enhanced features in each file type (QBO,OFX, CSV) that include:1. Each transaction in the QBO, OFX and CSV file is tagged with a unique identifier (a”hash”) so that duplication within the accounting system is no longer possible.2. Each QBO, OFX and CSV file, which are generated for each bank or credit card state-ment, is reconciled by the system and entries are placed into the CSV file for ”previ-ous balance,” ”ending balance” and ”reconciliation error” (if one is found) so thatwhen the file is uploaded into the accounting package, the user can easily see thestate of the statement reconciliation. This also ensures that all transactions are ac-counted for in the reconciliation process.3. The payees (vendors or names) associated with each transaction that is extractedfrom the pdf statement is normalized by the system so that the ”garbage” that isusually in a credit card or bank statement name field no longer needs to be removedmanually by the User and they can now ”use” that data without manual clean up.

[0040] Unlike traditional accounting systems where ’description’ to account mapping isperformed through manual assignment or rule creation, this invention provides ’payee’or ’name’ data that allows trivial vendor to account mapping at the accounting systemstage.

[0041] This enhanced vendor-transaction data plus accounting model data (chart of ac-counts, etc) allows for direct AI transformation to end reports such as income statements,balance sheets, expense reports, vendor reports, budgeting etc. This is a conceptual break-through in accounting systems as it allows for the accounting model to be considered a”lens” through which to view the underlying transaction data.

[0042] Furthermore, traditional accounting systems do not require vendor / payee but re-quire ’account’ or model attribute assignment. The information is essentially immutableand cannot be recategorized without reaccessing the source ’description’ data manually.Our systempreserves the original vendor-transaction relationship, allowing for dy-namic recategorization as needed, which is particularly valuable for businesses that needto analyze their data from multiple perspectives or adapt to changing accounting require-ments.

[0043] Our experience shows that as many as 80% of CPA clients are using QuickBooks orequivalent software primarily to help organize expense data and do not require full ac-counting models for yearly taxes and budget purposes. This invention provides a stream-lined pathway to organized financial data that meets these needs more efficiently thantraditional methods.MULTI-MODEL APPLICATION CAPABILITIES

[0044] Traditional accounting systems apply only one model (Chart of Accounts) to a com-pany or individual, and that model is completely subjective rather than based in absolutehistorical fact. This invention takes a fundamentally different approach by maintaining atransaction database with correctly identified vendor / payee and transfer information asan objective historical record that serves as the factual basis for all subsequent analysis.

[0045] The system enables multiple accounting models to be simultaneously applied to thesame underlying transaction data. This capability provides several key advantages:1. Generation of different financial reports for different stakeholders or purposes with-out requiring separate data processing pipelines2. Application of alternative classification schemes for specialized analysis, such as taxplanning, budgeting, or performance analysis3. Direct comparison of results under different accountingmethodologies, allowing forstrategic decision-making based on multiple perspectives4. Adaptation to changing reporting requirements without reprocessing transactiondata, significantly reducing the time and cost associated with regulatory changes

[0046] This approach recognizes that accounting models are interpretive lenses applied toobjective transaction data, rather than inherent qualities of the transactions themselves.By separating the objective transaction record from the subjective accounting models, theinvention enables a level of flexibility and analytical capability not available in traditionalaccounting systems.EXAMPLE OUTPUT

[0047] Figure 5 gives a truncated example of an Uber Pro Card statement which showsgenerated ’Payee’ and ’FitId’ fields that are not present in the original statement. Injectedinformation examples include [Beginning Balance], [EndingBalance], [EFT->1723], [Rec-onciliation Error] and other information that are not present in the original statement butare generated by the system.Figure 5: CSV of Extracted and Analyzed Uber Pro Card StatementTransaction Id,Job Id,Bank,Account,Date,Amount,Currency,Transaction Type,Payee,Description,Statement Date,Account Type,Bank Id,FitId7318,319,Uber Pro Card,1723,2024-08-03,-50.00,USD,Decrease,Cash App,"CASH APP*REDACTED, San Francisco, CA",2024-08-31,Checking,3000,2A2267B5195C60E7~2024-08-03~17320,319,Uber Pro Card,1723,2024-08-04,-53.54,USD,Decrease,Mint Mobile,"MINT MOBILE , 800683XXXX , CA",2024-08-31,Checking,3000,5C3E3ECB6EF722E0~2024-08-04~17319,319,Uber Pro Card,1723,2024-08-04,60.27,USD,Increase,Square,"SQ *ID - GYKOMRXXX, SAN FRANCISCO, CA",2024-08-31,Checking,3000,A52D0AFF074F8E4~2024-08-04~17322,319,Uber Pro Card,1723,2024-08-06,-50.00,USD,Decrease,Cash App,"CASH APP*REDACTED*ADD C, San Francisco, CA",2024-08-31,Checking,3000,42CB0444E671C614~2024-08-06~17321,319,Uber Pro Card,1723,2024-08-06,45.00,USD,Increase,[EFT->1723],"ACH Credit: EFT GEICO INSURANCE:10808:REDACTED, EFT",2024-08-31,Checking,3000,79E869B97A3F76B7~2024-08-06~17323,319,Uber Pro Card,1723,2024-08-18,70.96,USD,Increase,Square,"SQ *ID - 3LLN1Y87, SAN FRANCISCO, CA",2024-08-31,Checking,3000,11AAC23D2D6F6EDD~2024-08-18~17311,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Beginning balance],[ +50.00 ],2024-08-31,Checking,3000,7313,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Change in balance],[ -50.00 ],2024-08-31,Checking,3000,7312,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Ending balance],[ +0.00 ],2024-08-31,Checking,3000,7317,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Negative transactions],"[ -1,508.85 ]",2024-08-31,Checking,3000,7316,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Positive transactions],"[ +1,458.85 ]",2024-08-31,Checking,3000,7315,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Reconciliation error],[ +0.00 ],2024-08-31,Checking,3000,7314,319,Uber Pro Card,1723,2024-08-31,0.00,USD,OTHER,[Total transactions],[ -50.00 ],2024-08-31,Checking,3000,

[0048] Other files that are generated by the system include QBO,OFX and files for Quick-Books and other accounting systems that are subsets of this data and exactly compatiblewith the target accounting system.COMPARATIVE ANALYSIS WITH ALTERNATIVE TECHNOLOGIESTable 7: Comparison of PDF Processing vs. OFX Direct ConnectFeature OFX Direct Connect This Invention (PDF)Global Bank Support ~5% 100%US Bank Support ~15% 100%Checking Access 90 days 7+ yearsCredit Card Access 90 days 2+ yearsFitId Uniqueness No YesStatement Reconciliation No YesVendor Naming No YesNon-bank documents No Yes

[0049] Table 7 provides a direct comparison between the invention’s PDF-based approachand traditional OFX direct connect methods. The data demonstrates the technical andpractical advantages of the invention, addressing critical limitations of existing technolo-gies. CLAIMS WHEREAS Open Financial Exchange (OFX), QuickBooks Online (QBO), andcomma-separated values (CSV) data formats provided by financial insti-tutions do not include beginning and ending balance information for the

Claims

AMENDED CLAIMS received by the International Bureau on 20 January 2026 (20.01.2026)What is claimed:

1. A method for processing and analyzing financial data from PDF statements, comprising:(a) converting PDF statements into semi-structured text using a conversion module;(b) preprocessing the semi-structured text by injecting line numbers and tagging numerical data with character offset positions;(c) partitioning the preprocessed data into sections processable by a large language model;(d) analyzing the partitioned data using the large language model to extract transaction details;(e) reconciling the extracted transaction details with beginning and ending balances;(f) normalizing vendor information from transaction data; and(g) generating output files with injected vendor, reconciliation and transfer information in standardized formats.

2. The method of claim 1, wherein the preprocessing step enhances LLM analysis by:(a) enabling partitioning of data into context windows;(b) providing unique line identification;(c) allowing concurrent processing of partitioned data by multiple LLM instances;(d) facilitating error correction through feedback loop; and(e) enabling coherent reassembly of separately processed sections.

3. The method of claim 1, wherein the vendor normalization step uniquely identifies vendors from transaction and statement context, standardizing multiple variations to single normalized vendor names, wherein the vendor information is injected into payee fields.

4. The method of claim 1, wherein statement transactions are reconciled to beginning and ending balances, wherein the generated output files include reconciliation markers with beginning balance, ending balance, and reconciliation status entries.

5. The method of claim 1, further comprising processing directive files to:(a) group multiple vendors into single categories;(b) aggregate transactions from specified vendors;(c) apply vendor renaming and aggregation simultaneously; and(d) selectively redact vendor information.

6. The method of claim 1, wherein the output generation creates enhanced files not available from existing systems, achieving processing time reduction from 84.36 to 0.35 minutes.

7. The method of claim 1, further comprising:(a) storing normalized transaction data in a vendor database;(b) accepting Chart of Accounts mapping information;(c) directly generating financial reports using the transaction database;(d) bypassing traditional accounting system intermediary steps; and(e) providing capability to generate multiple financial reports from the same underlying data.

8. A method for vendor and payee name normalization in financial transaction data, comprising:(a) extracting raw transaction data from financial statements;(b) identifying variations in vendor name representations;(c) analyzing payee descriptions using a large language model;(d) creating standardized vendor name mappings;(e) applying the standardized mappings to normalize transaction data;(f) generating unique identifiers for each normalized transaction; and(g) maintaining a vendor database that learns new variations.

349. The method of claim 8, wherein the analysis uses contextual information including transaction amounts, dates, and surrounding transactions to improve vendor identification accuracy, achieving a 1:1 vendor-to-rule ratio.

10. A method for preprocessing semi-structured financial statement text, comprising:(a) injecting line numbers into the text creating structure [00001]- [99999] ;(b) tagging numerical data with character offset positions;(c) identifying column headers aligned with statement structure; and(d) outputting enhanced text for large language model analysis.

11. A system for processing and analyzing financial data from PDF statements, comprising:(a) a preprocessing module configured to convert PDF statements;(b) a partitioning module configured to divide preprocessed data;(c) an analysis module utilizing a large language model;(d) a reconciliation module configured to validate transactions;(e) a normalization module configured to standardize payee information;(f) an output module configured to generate standardized formats;(g) a directive processor configured to customize analysis; and(h) a vendor database for storing normalized vendor information.

12. The system of claim 11, wherein the preprocessing module injects line numbers and tags numerical data with character offset positions to improve large language model analysis, enabling data partitioning.

13. The system of claim 11, wherein the normalization module achieves 49% rule reduction by maintaining a 1:1 vendor-to-rule relationship.

14. The system of claim 11, wherein the output module generates QBO, OFX, and CSV formats with unique transaction identifiers and reconciliation information.

15. The system of claim 11, wherein the system achieves:(a) processing time reduction of 241x;(b) reconciliation error rate decrease from 1 in 5 to 1 in 100 statements;(c) ability to process over 50,000 transactions per hour;(d) perfect 1:1 vendor-to-rule relationship; and(e) cost reduction from $2.00 to $0.19 per transaction.

16. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 1.

17. The non-transitory computer- readable medium of claim 16, wherein the instructions further cause the processor to implement line numbering transformation and partitioning method.

18. A method for handling large financial statements exceeding LLM context limits, comprising:(a) detecting a statement exceeding context threshold;(b) partitioning the statement using line-indexed boundaries;(c) processing partitions in parallel using multiple LLM instances;(d) calculating subtotals for each partition;(e) assembling results with reconciliation verification; and(f) confirming mathematical accuracy of reassembly.

19. The method of claim 18, wherein each partition maintains line number continuity and position tags enabling column alignment preservation.

20. A method for generating enhanced financial output files, comprising:(a) creating unique Fitld identifiers in format [BANK] [ACCT]- [SEQ] ;(b) inserting reconciliation markers for beginning and ending balances;(c) normalizing all vendor names to standardized format;(d) adding statement boundary markers;(e) generating files in QBO, OFX, and CSV formats; and(f) validating reconciliation status for each statement period.35