Week 8 Worklog

Week 8 Objectives:

• Study the business problem and MVP scope for DocuFlow AI - a serverless intelligent invoice and receipt processing platform on AWS.
• Identify the required AWS services, operating cost, technical risks, and high-level solution architecture.
• Prepare the initial project proposal as the baseline for the workshop and module ownership.

Tasks to be carried out this week:

Week 8 Achievements:

• Understood the business problem of DocuFlow AI: reducing manual data entry for invoices and receipts, improving status tracking, normalizing extracted data, and supporting auditability for Finance/Ops teams.

• Locked the MVP scope to a workshop-friendly end-to-end flow:

  • Users authenticate with Cognito.
  • Users upload PDF/JPG/PNG files through the frontend and a presigned URL.
  • The S3 raw bucket triggers an asynchronous processing workflow.
  • Textract extracts invoice/receipt data.
  • Bedrock normalizes, classifies, and returns output using a shared JSON schema.
  • DynamoDB stores metadata/status, and S3 processed stores JSON/CSV results.
  • CloudWatch and SNS/SES notify failures or low-confidence documents.

• Identified the main AWS services and their roles in the serverless, event-driven architecture:

  • Cognito for EndUser, Reviewer, and Admin authentication/authorization.
  • API Gateway + Lambda for presigned URL creation, status updates, and lightweight backend logic.
  • S3 for raw documents and processed results.
  • EventBridge + SQS + Step Functions for asynchronous processing, retry, and execution audit.
  • Textract + Bedrock for extraction, normalization, classification, and schema-based output.
  • DynamoDB for the status table and metadata.
  • CloudWatch + SNS/SES for logs, metrics, alarms, and notifications.
  • IAM/KMS for least privilege and data encryption.

• Analyzed cost and proposed control measures:

  • Limit the demo to 5-10 small invoice/receipt files.
  • Track services billed by request, page, or token, especially Textract, Bedrock, and Step Functions.
  • Query only the processed demo folder when using Athena to avoid unnecessary scan cost.
  • Configure S3 lifecycle policies for raw/processed data and prepare cleanup/destroy commands after the demo.
  • Set up budget alerts to catch unexpected cost.

• Built the initial risk matrix and mitigation plan:

  • Oversized scope: keep the MVP focused on invoice/receipt processing and move contracts, purchase orders, email ingestion, and advanced QuickSight dashboards to extensions.
  • OCR errors on low-quality documents: use confidence scores, a REVIEW_REQUIRED threshold, and a manual review loop.
  • Bedrock schema mismatch: define strict prompts/schema, validate output with Lambda, and add retry/fallback behavior.
  • Overly broad IAM: create module-specific roles, apply least privilege, and avoid hard-coded access keys.
  • Unexpected AI cost: limit demo files, enable budget alerts, and clean up resources after the workshop.

• Completed the first end-to-end architecture draft for DocuFlow AI, covering client/access, ingestion/workflow, AI processing, storage, analytics, observability, and DevOps/IaC layers.

• Completed the initial project proposal with the following sections:

  • Executive summary and problem statement.
  • MVP scope, out-of-scope items, and success criteria.
  • Solution architecture and service selection.
  • End-to-end workflow.
  • Data contract, status model, and five-person module ownership.
  • Security baseline, observability, analytics, cost control, risk mitigation, and definition of done.