Has anyone used Relevance AI for data cleaning pipelines? Like standardizing CSV data?

I've been conducting a series of evaluations on various AI platforms for data preprocessing tasks, specifically focusing on the automation of cleaning and standardizing heterogeneous CSV files. The promise of platforms like Relevance AI is to move beyond simple prompt-chaining and into robust, production-ready pipelines. My question stems from a practical experiment: can it reliably handle the messy reality of real-world tabular data?

In my testing, I've used several other AI workflow tools for similar tasks, often running into issues with consistent parsing of date formats, categorical value normalization, and handling of missing values across large batches. Relevance AI's agent-based architecture and the concept of "tools" seems theoretically well-suited for this. I'm particularly interested in whether its data-specific tools (like those for parsing, transformation, and validation) can be orchestrated effectively for a pipeline that does the following:

* Ingests a CSV with inconsistent column naming (e.g., "First_Name," "first name," "fname").
* Standardizes data types (dates, currencies, percentages) from multiple regional formats.
* Maps categorical values to a controlled vocabulary (e.g., "USA," "U.S.A.," "United States" -> "US").
* Outputs a cleaned dataset, preferably with a validation report.

I attempted to prototype this using their Studio interface. The visual workflow builder is clear, but the devil is in the details of configuration. For instance, when using a `transform_data` tool, the prompt configuration for field-specific rules becomes critical. Here's a simplified example of the kind of agent configuration I'm exploring:

```yaml
task: "Standardize and clean the incoming customer data CSV."
tools:
- csv_parser:
delimiter: "auto"
encoding: "utf-8"
- transform_data:
instructions: |
- For column 'date': Parse to ISO 8601 (YYYY-MM-DD). Assume MM/DD/YYYY if ambiguous.
- For column 'amount': Extract numeric value, remove currency symbols. Standardize to float.
- For column 'country': Apply mapping to 2-letter ISO codes. Log unmappable values.
- validate_dataset:
checks:
- no_empty_rows: true
- allowed_categories:
column: 'status'
values: ['active', 'inactive', 'pending']
```

My primary concerns are around scalability and deterministic results. When processing, say, 10,000 rows with such a pipeline:
* Does the LLM-powered transformation introduce non-determinism in outputs, or can it be sufficiently constrained?
* What are the practical costs for such a volume, given the token consumption per row operation?
* How does error handling work? If one row fails parsing, does the entire batch halt, or is there a recovery mechanism?

I would be very interested to hear from anyone who has moved beyond a demo and used Relevance AI for a sustained data cleaning operation. Comparative insights on its performance versus writing custom Python scripts with Pandas/Polars, or using other specialized ETL platforms, would be particularly valuable for my benchmarking framework.