Granted Patent B1

US 12,737,374 · App. 19/568,539 · Granted Sep 15, 2026

Inventor: Andrew Jagoe (Castle Pines, CO)

Assignee: Minted LLC

G06F16/252G06F16/2282G06F40/205G06F40/295

Abstract

Claims (112)

1 . A method implemented by one or more computing devices, the method comprising:

receiving an input file containing contact information;

parsing the input file to extract columns and rows of contact data;

computing an essential-field coverage metric based on the parsed input file, the essential-field coverage metric quantifying a degree to which parsed columns map to required contact fields;

responsive to the essential-field coverage metric meeting or exceeding a threshold, performing rule-based column mapping using exact header matching against one or more known external formats to generate a first mapping;

responsive to the essential-field coverage metric being below the threshold or rule-based column mapping being unavailable, invoking a large language model-based column mapping engine to produce a second mapping, the second mapping comprising a structured mapping output conforming to a predefined schema;

extracting contact information from the input file using the first mapping or the second mapping to generate structured contact data,

verifying an address within the structured contact data, the verifying comprising;

classifying the address as an address of a first type, an address of a second type, or an address of an unknown type;

responsive to the address being classified as the first type, calling a first verification service to confirm accuracy of the address;

responsive to the address being classified as the second type, calling a second verification service to confirm accuracy of the address, the second verification service being different from the first verification service; and

responsive to the address being classified as the unknown type, invoking a large language model to interpret address components and confirm accuracy of the address; and

providing the structured contact data to a consuming application or process.

2 . The method of claim 1 , further comprising:

automatically detecting a file format of the input file by inspecting file extensions to identify format type, the file format being one of a delimited text format, a spreadsheet workbook format, or an electronic business card format.

3 . The method of claim 2 , further comprising:

responsive to the file format being the delimited text format, tokenizing records by line boundaries and splitting field values by a detected delimiter character with handling for quoted field values and escape sequences.

4 . The method of claim 2 , further comprising:

responsive to the file format being the spreadsheet workbook format, traversing cell ranges within one or more worksheets and extracting cell values along with data-type annotations.

5 . The method of claim 2 , further comprising:

responsive to the file format being the electronic business card format, interpreting property-value pairs according to a predefined specification and mapping the property-value pairs to a columnar representation.

6 . The method of claim 1 , further comprising:

generating metadata comprising a mapping confidence score, the essential-field coverage metric, and a decision reasoning record that documents whether the first mapping or the second mapping was selected, the metadata being serialized alongside the structured contact data for downstream audit and quality assessment.

7 . The method of claim 1 , further comprising:

responsive to detecting a failure condition comprising at least one of a network timeout, an inference endpoint outage, or an inference failure when invoking the large language model-based column mapping engine, reverting to a fallback mapping using relaxed heuristics.

8 . The method of claim 7 , wherein the relaxed heuristics comprise at least one of: normalization-based matching, underscore-to-space conversion, and abbreviation expansion; or fuzzy header matching using Levenshtein distance, Jaro-Winkler similarity, or substring containment.

9 . The method of claim 1 , further comprising:

normalizing the structured contact data by converting street type abbreviations to standardized forms, normalizing state and province names to standard postal abbreviations, formatting postal codes according to locale-specific patterns, and standardizing country names to a predefined specification.

10 . The method of claim 1 , further comprising:

validating the structured contact data by verifying that required fields are present, verifying that field values conform to expected data types and formats, verifying that field lengths do not exceed maximum permitted values, and annotating rejected contact rows with failure codes for remediation.

11 . The method of claim 1 , further comprising:

selecting among a plurality of inference service endpoints based on cost-optimized decision making that considers inference application programming interface (API) call latency, monetary cost per API call, current queue depth, and historical performance data.

12 . The method of claim 1 , further comprising:

prior to computing the essential-field coverage metric, performing smart header row detection comprising:

evaluating a first row of the input file as a primary header candidate;

responsive to validation of the first row failing, detecting exception patterns comprising at least one of empty rows, instruction blocks, metadata rows, or title blocks;

searching for an actual header row using characteristic analysis comprising matching cell contents against known header terms and evaluating keyword density; and

responsive to no header row being identified through characteristic analysis, falling back to treating the first row as the header row.

13 . The method of claim 1 , wherein performing rule-based column mapping comprises:

in a first phase, performing exact string matching of source column headers against a header lexicon; and

in a second phase, for headers not matched in the first phase, performing normalized matching comprising case normalization, underscore-to-space conversion, and abbreviation expansion.

14 . The method of claim 1 , wherein the threshold comprises a dual-threshold configuration, the method further comprising:

responsive to the essential-field coverage metric indicating that essential fields are satisfied, applying a lower confidence threshold for accepting the first mapping; and

responsive to the essential-field coverage metric indicating that essential fields are not satisfied, applying a higher confidence threshold for accepting the first mapping.

15 . The method of claim 1 , wherein providing the structured contact data to the consuming application or process comprises transmitting, via a network interface, the structured contact data as a serialized data payload to a remote server, thereby reducing network bandwidth consumption compared to transmitting the input file in its original format.

16 . The method of claim 1 , further comprising:

instantiating, by one or more hardware processors, a file processor in a memory buffer, the file processor configured to receive the input file via a network interface;

generating, by the file processor, a parsed data structure stored in the memory buffer, the parsed data structure comprising a two-dimensional array of column headers and row values; and

transmitting the parsed data structure to a column mapper for the rule-based column mapping or the large language model-based column mapping.

17 . The method of claim 1 , further comprising:

scaling, by a container orchestration system, a file processor, a column mapper, and a contact extractor independently based on processing load metrics; and

allocating additional processing threads or memory buffers to bottleneck stages without over-provisioning resources for stages operating within capacity,

wherein the scaling enables targeted resource allocation that conserves CPU cycles, memory allocations, and network bandwidth compared to non-scaled processing.

18 . A computer-implemented method for ingesting and normalizing contact data using a modular multi-stage processing pipeline executing on a server computing device, the method comprising:

instantiating, by one or more hardware processors of the server computing device, a file processor in a memory buffer, the file processor configured to receive an input file via a network interface and automatically detect a file format of the input file by inspecting file extensions;

parsing, by the file processor executing on the one or more hardware processors, the input file according to the detected file format to generate a parsed data structure stored in the memory buffer, the parsed data structure comprising a two-dimensional array of column headers and row values;

instantiating, by the one or more hardware processors, a column mapper comprising a rule-based mapping engine and a large language model-powered mapping engine;

executing, by the rule-based mapping engine, exact header matching against a header lexicon to generate a first mapping associating source column indices with canonical field identifiers, and concurrently computing an essential-field coverage metric quantifying presence of name and address fields;

responsive to the essential-field coverage metric being below a configurable threshold, invoking, by the large language model-powered mapping engine, an inference service endpoint by transmitting a prompt payload comprising the column headers and a description of a target canonical schema via the network interface, and receiving a structured mapping output constrained to a predefined output schema encoded as a JSON schema or function-calling specification;

instantiating, by the one or more hardware processors, a contact extractor configured to receive the first mapping or the structured mapping output and the parsed data structure;

extracting, by the contact extractor, contact field values from each row of the parsed data structure using the received mapping, and normalizing the extracted field values into canonicalized contact records conforming to the target canonical schema;

validating, by the contact extractor, the canonicalized contact records by performing field-level checks comprising data type verification, format verification, and length verification, and annotating rejected rows with failure codes;

instantiating, by the one or more hardware processors, a metadata generator configured to generate metadata comprising mapping confidence, essential-field coverage percentage, and decision reasoning; and

outputting, by the one or more hardware processors, a canonicalized contact dataset comprising the canonicalized contact records and the metadata to a downstream consuming application via the network interface.

19 . The method of claim 18 , further comprising:

executing, by the column mapper, a defined fallback sequence comprising:

attempting rule-based mapping using the rule-based mapping engine,

evaluating the essential-field coverage metric,

invoking the large language model-powered mapping engine when the essential-field coverage metric is below the configurable threshold, and

reverting to rule-based fallback with relaxed heuristics upon detection of inference endpoint unavailability or failure.

20 . The method of claim 19 , further comprising:

recording, by the metadata generator, fallback events comprising a timestamp, an identifier of the inference service endpoint that failed, and an error code, the fallback events being included in the metadata for downstream audit.

21 . The method of claim 18 , further comprising:

dynamically selecting, by the one or more hardware processors, among a plurality of inference service endpoints operated by different large language model providers, the dynamic selection being based on service availability indicators and latency measurements obtained via health check requests transmitted via the network interface.

22 . The method of claim 18 , further comprising:

adapting, by the one or more hardware processors, cost-optimized decision making over time by logging provider selection events and applying a reinforcement learning algorithm or multi-armed bandit algorithm that balances exploration of provider options with exploitation of historically low-cost, high-reliability providers.

23 . The method of claim 18 , further comprising:

scaling, by a container orchestration system, the file processor, the column mapper, and the contact extractor independently based on processing load metrics, thereby allocating additional processing threads or memory buffers to bottleneck stages without over-provisioning resources for stages operating within capacity.

24 . The method of claim 18 , further comprising:

receiving, by the large language model-powered mapping engine in the structured mapping output, a confidence score per mapped field, and an explanation token sequence generated by the inference service endpoint articulating reasoning behind each mapping decision.

25 . A system comprising:

one or more processors; and

memory storing instructions that, when executed by the one or more processors, cause the system to:

receive an input file containing contact information;

parse the input file to extract columns and rows of contact data;

compute an essential-field coverage metric based on the parsed input file, the essential-field coverage metric quantifying a degree to which parsed columns map to required contact fields;

responsive to the essential-field coverage metric meeting or exceeding a threshold, perform rule-based column mapping using exact header matching against one or more known external formats to generate a first mapping;

responsive to the essential-field coverage metric being below the threshold or rule-based column mapping being unavailable, invoke a large language model-based column mapping engine to produce a second mapping, the second mapping comprising a structured mapping output conforming to a predefined schema;

extract contact information from the input file using the first mapping or the second mapping to generate structured contact data;

verify an address within the structured contact data by:

classifying the address as an address of a first type, an address of a second type, or an address of an unknown type;

responsive to the address being classified as the first type, calling a first verification service to confirm accuracy of the address;

responsive to the address being classified as the second type, calling a second verification service to confirm accuracy of the address, the second verification service being different from the first verification service; and

responsive to the address being classified as the unknown type, invoking a large language model to interpret address components and confirm accuracy of the address; and

provide the structured contact data to a consuming application or process.

26 . The system of claim 25 , wherein the instructions further cause the system to: automatically detect a file format of the input file by inspecting file extensions to identify format type, the file format being one of a delimited text format, a spreadsheet workbook format, or an electronic business card format.

27 . The system of claim 25 , wherein the instructions further cause the system to: generate metadata comprising a mapping confidence score, the essential-field coverage metric, and a decision reasoning record that documents whether the first mapping or the second mapping was selected, the metadata being serialized alongside the structured contact data for downstream audit and quality assessment.

28 . The system of claim 25 , wherein the instructions further cause the system to: responsive to detecting a failure condition comprising at least one of a network timeout, an inference endpoint outage, or an inference failure when invoking the large language model-based column mapping engine, revert to a fallback mapping using relaxed heuristics.

29 . The system of claim 25 , wherein the instructions further cause the system to: normalize the structured contact data by converting street type abbreviations to standardized forms, normalizing state and province names to standard postal abbreviations, formatting postal codes according to locale-specific patterns, and standardizing country names to a predefined specification.

30 . A system for ingesting and normalizing contact data using a modular multi-stage processing pipeline, the system comprising:

one or more hardware processors of a server computing device; and

memory storing instructions that, when executed by the one or more hardware processors, cause the system to:

instantiate a file processor in a memory buffer, the file processor configured to receive an input file via a network interface and automatically detect a file format of the input file by inspecting file extensions;

parse, by the file processor, the input file according to the detected file format to generate a parsed data structure stored in the memory buffer, the parsed data structure comprising a two-dimensional array of column headers and row values;

instantiate a column mapper comprising a rule-based mapping engine and a large language model-powered mapping engine;

execute, by the rule-based mapping engine, exact header matching against a header lexicon to generate a first mapping associating source column indices with canonical field identifiers, and concurrently compute an essential-field coverage metric quantifying presence of name and address fields;

responsive to the essential-field coverage metric being below a configurable threshold, invoke, by the large language model-powered mapping engine, an inference service endpoint by transmitting a prompt payload comprising the column headers and a description of a target canonical schema via the network interface, and receive a structured mapping output constrained to a predefined output schema encoded as a JSON schema or function-calling specification;

instantiate a contact extractor configured to receive the first mapping or the structured mapping output and the parsed data structure;

extract, by the contact extractor, contact field values from each row of the parsed data structure using the received mapping, and normalize the extracted field values into canonicalized contact records conforming to the target canonical schema;

validate, by the contact extractor, the canonicalized contact records by performing field-level checks comprising data type verification, format verification, and length verification, and annotate rejected rows with failure codes;

instantiate a metadata generator configured to generate metadata comprising mapping confidence, essential-field coverage percentage, and decision reasoning; and

output a canonicalized contact dataset comprising the canonicalized contact records and the metadata to a downstream consuming application via the network interface.

Assignments (1)

References Cited (14)