Machines / DP-04
MODEL DP-04
Data pipeline
Messy sources in. Clean, ranked records out.
LLM extraction, cleanup and ranking over résumés, PDFs, profiles and web pages — into structured, searchable records with schema checks and sampled human audits.
In plain words: Invoice, résumé and document extraction.
INPUTRésumés, PDFs, profiles, web pages
OUTPUTClean, ranked, searchable records
RUNS ONBatch or streaming on your infra
OPERATORA person audits a sample and handles flagged items
Operating manual
Rev. 1 · DP-04§1
What it does
Reads each document, extracts the fields you define, validates them against a schema, de-duplicates, and ranks or routes the result. Anything that fails validation is flagged for a person instead of being silently guessed.
§2
Install when
- The same fields are typed out of documents every week
- Volume is too high to read everything
- Downstream systems need clean, consistent records
§3
Do not install when
- Documents are mostly unreadable scans with no OCR budget
- The fields change every month
- No one can define what "correct" looks like
§4 · The line
Each station is a separate step with one job and its own check. A failure is caught where it happens, not at the end.
01IntakeFiles, URLs, email
02ReadOCR + parsing
03ExtractFields by schema
04ValidateTypes, ranges, duplicates
05Rank + storeSearchable records
↺ Validation fails → flagged for a person · a sample of passes is audited every run
§5 · Safety features
Schema or nothingA record that doesn't validate is never stored as fact.
Human sampleA fixed share of every batch is checked by a person.
Source keptEvery field links back to where it came from.
PII handlingMasking and retention rules set with you.
§6 · Gauges
What's watched on every install. Calibrated per team.Field accuracyOn the audited sample
Flag rateShare needing a person
ThroughputDocuments per hour
Cost per documentCompute per record
§7 · Parts bin
OCR / parser
LLM extractor
Schema validator
De-duplicator
Search index