Most document processing solutions cost hundreds per month and take weeks to set up. This one costs $12/mo at 1,000 documents and took a weekend to build.
Here's the full breakdown — architecture, stack, costs, and how to deploy it yourself.
The Problem
Every business has a pile of PDFs that need to become structured data. Invoices, contracts, reports, applications. Someone manually reads each one, copies values into a spreadsheet, and hopes they don't make mistakes.
That someone's time is worth a lot more than $12/month.
Architecture Overview
The pipeline is simple — three AWS services and one API call:
- ·S3 Bucket — PDFs land here (uploaded by users or another system)
- ·Lambda Function — Triggered by S3 events, reads the PDF, sends to Claude API
- ·Claude API — Extracts structured data from the document content
- ·PostgreSQL (RDS) — Stores the extracted data in a normalized schema
No servers to manage. No containers. No orchestration. Just events flowing through serverless functions.
The Stack
| Component | Service | Cost at 1K docs/mo | | ------------- | ------------------ | ------------------ | | Storage | S3 | ~$0.02 | | Compute | Lambda | ~$0.50 | | AI Extraction | Claude API (Haiku) | ~$8.00 | | Database | RDS (db.t3.micro) | ~$3.50 | | Total | | ~$12.02 |
The biggest cost is the Claude API — and even that's cheap because we use Claude Haiku for extraction. Haiku is fast, cheap, and more than capable of pulling structured fields from documents.
How It Works
Step 1: PDF Upload Triggers Lambda
When a PDF lands in the S3 bucket, an S3 event notification triggers the Lambda function. No polling, no cron jobs — it's instant.
def handler(event, context):
bucket = event['Records'][0]['s3']['bucket']['name']
key = event['Records'][0]['s3']['object']['key']
# Download the PDF from S3
pdf_bytes = s3.get_object(Bucket=bucket, Key=key)['Body'].read()
Step 2: Extract Text from PDF
We use pymupdf (lightweight, Lambda-friendly) to extract raw text:
import pymupdf
doc = pymupdf.open(stream=pdf_bytes, filetype="pdf")
text = "\n".join(page.get_text() for page in doc)
Step 3: Claude API Extracts Structured Data
This is where the magic happens. We send the raw text to Claude with a structured extraction prompt:
response = anthropic.messages.create(
model="claude-haiku-4-5-20251001",
max_tokens=1024,
messages=[{
"role": "user",
"content": f"""Extract the following fields from this document:
- vendor_name
- invoice_number
- date
- line_items (array of: description, quantity, unit_price, total)
- subtotal
- tax
- total
Return as JSON only, no explanation.
Document:
{text}"""
}]
)
data = json.loads(response.content[0].text)
Step 4: Store in PostgreSQL
The extracted JSON goes straight into a normalized database:
cursor.execute("""
INSERT INTO documents (s3_key, vendor, invoice_number, date, total, raw_data)
VALUES (%s, %s, %s, %s, %s, %s)
""", (key, data['vendor_name'], data['invoice_number'],
data['date'], data['total'], json.dumps(data)))
Cost Breakdown at Scale
| Documents/mo | Claude API | Lambda | S3 | RDS | Total | | ------------ | ---------- | ------ | ----- | ------ | ----------- | | 100 | $0.80 | $0.05 | $0.01 | $3.50 | $4.36 | | 1,000 | $8.00 | $0.50 | $0.02 | $3.50 | $12.02 | | 10,000 | $80.00 | $5.00 | $0.23 | $15.00 | $100.23 |
At 10K documents/month you're paying about 1 cent per document. Try getting a human to process a document for a cent.
Deployment
The entire infrastructure is defined in AWS CDK — one cdk deploy and you're live:
git clone https://github.com/Crisfon6-dev/doc-processor-template
cd doc-processor-template
npm install
cdk deploy --all
The CDK stack creates: S3 bucket with event notifications, Lambda function with pymupdf layer, RDS instance, IAM roles, and VPC networking. Zero manual AWS console clicks.
What I Learned
Claude Haiku is surprisingly good at extraction. I started with Sonnet thinking I'd need the extra capability, but Haiku handles structured extraction perfectly — and at 1/10th the cost.
S3 event triggers are the right pattern here. I considered SQS, Step Functions, and EventBridge. All overkill. S3 → Lambda is the simplest possible pipeline and it works.
The hardest part was PDF text extraction, not AI. Some PDFs have terrible text layers. pymupdf handles most cases, but scanned documents need OCR (Textract) which adds cost and complexity.
Get the Template
The full template — CDK stack, Lambda function, extraction prompts, and database schema — is available on my GitHub. Clone it, configure your Anthropic API key, deploy.
This is template #1 in the PowerAI weekly series. Every week I publish a new production-ready automation with architecture diagrams, cost breakdowns, and working code.