TECHNIQUE
Computer Vision
OCR & document vision is being used as an intake/inspection layer for invoices, education submissions, fulfillment quality, benefits forms, and eKYC documents, with operators differing mainly on managed IDP vs internal/custom model stacks and how much human review remains.
Use OCR/CV to machine-read operational documents or product images before downstream handling: Uber extracts text from document images, BQA uses Amazon Textract to extract document text, Amazon uses OCR/CV on fulfillment-center images, and Missouri automates extraction from paper forms.
4 of 4 deployed/pilot operators.Combine document vision/OCR with language or multimodal models for interpretation beyond raw text: Uber uses LLMs to extract invoice fields, BQA sends extracted text to summarization and assessment models, Amazon uses an MLLM to describe product damage, and Grab announced OCR/KIE with a Vision LLM stack.
3 of 4 deployed/pilot operators; Grab is announced and is not counted in the deployed/pilot arithmetic.Keep humans in the loop for critical review or final determinations: Uber designed HITL review and correction with side-by-side PDF/model output comparison; Missouri states final eligibility determinations are always made by trained DSS staff.
2 of 4 deployed/pilot operators.Stage documents through ingestion, storage, queues, or preprocessing before extraction: Uber ingests emails/PDFs/tickets into object storage and standardizes document formats; BQA stores uploads in S3, queues them with SQS, and invokes Lambda/Textract extraction.
2 of 4 deployed/pilot operators.Train, fine-tune, or evaluate models on task/domain data: Uber fine-tuned and evaluated multiple LLMs for invoice processing; Amazon trains CV models with catalog reference images and actual product images; Grab announced full-parameter fine-tuning and task-centric fine-tuning for document data.
2 of 4 deployed/pilot operators; Grab is announced and is not counted in the deployed/pilot arithmetic.Instrument document processing with operational metrics: Uber captures processing speed, accuracy rates, and cost efficiency for its document platform.
1 of 4 deployed/pilot operators.Across the deployed/pilot pool, OCR/document vision feeds a downstream workflow rather than ending at text extraction: Uber sends approved invoices to ERP, BQA stores summaries and assessment results, Amazon triggers defect investigation/removal steps, and Missouri routes extracted form data to trained staff for eligibility determinations.
Every deployed/pilot operator applied the technique to a manual, high-volume, or error-prone operational bottleneck: invoice extraction delays at Uber, manual submission review at BQA, imperfect-product detection at Amazon, and increased benefits paperwork/manual processing at Missouri.
Architecture choices differ substantially by operator.
APPROACH 01
Managed AWS-style IDP/event pipeline with S3, SQS, Lambda, Textract, SageMaker/Bedrock models.
APPROACH 02
Internal reusable document platform plus internal CV/OCR gateway for invoices and other document types.
APPROACH 03
Fulfillment-center imaging tunnels with OCR, CV models, MLLM, and ensemble models for product-quality inspection.
APPROACH 04
Packaged document automation for SNAP forms, with automation handling data extraction and trained staff making final eligibility decisions.
APPROACH 05
Announced custom lightweight Vision LLM and document-understanding platform with detection, orientation, OCR, and KIE modules.
Human involvement is placed at different parts of the lifecycle.
APPROACH 01
Production review or final decision authority after automated extraction.
APPROACH 02
Human review to improve label accuracy during dataset/model preparation in the announced document-vision build.
The output expected from OCR/document vision differs by workflow.
APPROACH 01
Structured invoice fields and ERP/payment handoff.
APPROACH 02
Summaries, comparisons, assessments, and generated comments for education submissions.
APPROACH 03
Expiration/label checks, damage detection, and plain-language defect reports.
APPROACH 04
Initial extraction from eligibility paperwork, with eligibility decisions reserved for trained staff.
APPROACH 05
JSON key-information extraction from unstructured OCR text in the announced eKYC/document-processing workflow.
Hard-to-read documents remain a practical constraint: Uber explicitly preprocesses low-resolution scans and handwritten text, while Missouri reports historical struggles with incomplete or illegible documents.
Document diversity is a recurring source of complexity: Uber cites thousands of suppliers, varying invoice templates, and more than 25 languages; Missouri cites over 30 eligibility document types across hundreds of thousands of applicants; Grab’s announced work cites Southeast Asian language and document-format diversity.
Operators retained human control for consequential outputs: Uber uses HITL validation for critical reviews and corrections, and Missouri keeps final eligibility determinations with trained DSS staff.
| Name | Kind | When | Maturity |
|---|---|---|---|
| PaddleOCR | library | self-hosted OCR with strong multilingual accuracy and layout models | established |
| Tesseract | library | clean printed text on a budget: no GPU, permissive license, known limits | commodity |
| Azure Document Intelligence | service | managed forms and layout extraction with prebuilt invoice/receipt/ID models | established |