AI
CompletedAI-Powered Business Automation
Practical use of the Anthropic Claude and OpenAI APIs on repetitive tasks: reading documents, extracting data and preparing work for a person to approve.
Person approves5/5
A human makes the final decision.
The problem
Many operational tasks are repetitive but not simple. They involve reading a document, understanding it and deciding what happens next.
Full automation is not acceptable where a mistake has a cost, so the useful question is which step AI can take and which step a person must keep.
My contribution
- Built Python and TypeScript integrations with the Anthropic Claude and OpenAI APIs.
- Applied AI-assisted document processing to tasks such as extracting action items from notes.
- Explored batch processing for repeated document tasks.
- Built agents for recurring operational work.
- Designed provider-independent setups, so that a model can be replaced through configuration.
Technical architecture
- Models
- Hosted and local language models
- Orchestration
- Python scripts and TypeScript workers
- Guardrails
- Deterministic validation and human approval
- Output
- Drafts, structured data and reports for review
Technologies
- Claude API
- OpenAI API
- Python
- TypeScript
Highlights
- Human review built into every workflow
- Models can be replaced through configuration
Details are generalized to protect confidential information. Company systems, customers, figures, source code and screens are not shown.