Design NL-to-Formula assistant for Airtable

Scenario

You are given:

• An Airtable API key and a link/base/table you can read/write.
• An LLM API key (e.g., Claude) that you can call.

Users type natural-language requests such as:

• “add 2 to salary”
• “if status is 'active', set bonus to 10% of salary”
• “create a column that is salary * 1.1”

Your task is to build a service that:

1. Interprets the user request and the table schema.
2. Converts the request into an Airtable formula expression (or equivalent transformation) that can be applied to a field/column.
3. Applies the transformation to the online sheet via the Airtable API.
4. Returns a clear result to the user (what was changed, preview, and any errors).

Requirements

• Must handle ambiguous requests (ask clarifying questions when needed).
• Must be safe: prevent destructive/unsafe operations, data leakage, and prompt injection.
• Must be reliable: validate formulas before applying, handle API errors/rate limits, and support retries.
• Should be observable: logs/metrics/traces for debugging and evaluation.

Deliverables

• High-level architecture (components and data flow).
• Prompting/tool-calling strategy to translate natural language to an Airtable formula.
• Validation and safety checks before writing back to Airtable.
• How you would test and evaluate quality (offline + online).
• Key edge cases and failure modes.