AWS PDF Analyzer

Serverless pipeline that takes a PDF, extracts the text with AWS Textract, then sends it to OpenAI to pull out structured data — names (with IINs), decisions, and reasons. Everything runs on Lambda + Step Functions, results land in S3.

How it works

POST /documents
      ↓
  api_handler  →  S3 (raw PDF)  →  Step Functions
                                        ↓
                                   parse_lambda  (Textract)
                                        ↓
                                  analyze_lambda  (OpenAI)
                                        ↓
                                   S3 (result JSON)

Status is tracked in DynamoDB throughout. Once complete, the result JSON has:

• fullNames — array of { name, iin } found in the document
• decision — any conclusion or resolution
• reason — justification for the decision

Data layout

S3 — everything lives under documents/{doc_id}/:

Key	What
documents/{doc_id}/source.pdf	original uploaded PDF
documents/{doc_id}/intermediate/extracted_text.txt	raw text from Textract
documents/{doc_id}/result/final_output.json	final structured output

DynamoDB — one item per document, pk = DOC#{doc_id}, sk = META:

Field	Description
doc_id	UUID generated on upload
status	RECEIVED → PARSING → PARSED → ANALYZING → COMPLETED (or FAILED)
source_s3_key	path to the uploaded PDF in S3
text_s3_key	path to extracted text (set after Textract finishes)
result_s3_key	path to the result JSON (set after OpenAI finishes)
created_at / updated_at	ISO 8601 timestamps
error_message	only present on FAILED, capped at 500 chars

result_url in the GET response is a presigned S3 URL generated on the fly (valid 2 hours), not stored in DynamoDB.

Setup

1. Create a DynamoDB table (pk + sk as keys)
2. Create an S3 bucket
3. Create an empty Step Function
4. Create api_handler lambda, upload the zip, set env vars:
   • DOCUMENTS_BUCKET, DOCUMENTS_TABLE_NAME, STATE_MACHINE_ARN
5. Create parse_lambda lambda, upload the zip, set env vars:
   • DOCUMENTS_TABLE_NAME
6. Create analyze_lambda lambda, upload the zip, set env vars:
   • DOCUMENTS_TABLE_NAME, OPENAI_API_KEY, OPENAI_MODEL (default: gpt-4o-mini)
7. Attach IAM roles from /permissions to each lambda
8. Upload /statemachine/pipeline.asl.json to the Step Function, fill in the lambda ARNs in .parameters.FunctionName
9. Create an API Gateway with POST /documents and GET /documents/{doc_id}, both pointing to api_handler

Test

Upload a document:

curl -X POST https://<id>.execute-api.<region>.amazonaws.com/documents \
    -H "Content-Type: application/pdf" \
    --data-binary @test_pdfs/loan_request_seitkali.pdf

Response:

{"doc_id": "...", "message": "processing started", "execution_arn": "..."}

Check status:

curl https://<id>.execute-api.<region>.amazonaws.com/documents/<doc_id>

Response when done:

{
  "doc_id": "...",
  "status": "COMPLETED",
  "created_at": "2026-04-22T13:41:40.752680+00:00",
  "updated_at": "2026-04-22T13:41:58.553909+00:00",
  "result_url": "...",
  "result_s3_key": "documents/<doc_id>/result/final_output.json"
}

Open result_url to see the extracted data. There are a few test PDFs in /test_pdfs to try out.