Design an NL-to-formula assistant

ML system design: Natural-language to spreadsheet formula assistant

Design an assistant that converts natural language requests into spreadsheet-style formulas for a no-code table product (similar to Airtable/Sheets).

Users and examples

Users type requests like:

• “If Status is "Won" and Amount > 1000, return Amount * 0.1 else 0.”
• “Extract the domain from the Email field.”
• “Create a formula to label rows as "Late" if Due Date is before today and Completed is false.”

The system should output:

1. A proposed formula (in the product’s formula language),
2. A short explanation , and
3. Optionally a few test cases / example evaluations .

Requirements

• Support ambiguous requests by asking clarifying questions .
• Respect table schema/context (field names, types, missing values).
• Be safe: avoid leaking data, and avoid destructive actions.
• Latency target: e.g., P95 < 2–3s for interactive usage.
• Provide a plan for quality evaluation and iteration .

What to cover

• End-to-end architecture (client, orchestration, model calls, tools).
• Schema grounding and type checking.
• Handling ambiguity and multi-turn refinement.
• How you validate formulas (static + dynamic checks).
• Offline/online evaluation metrics and datasets.
• Failure modes and mitigations (hallucinated fields/functions, incorrect edge-case logic).