Anyone actually using Freeplay in production with 100k daily calls

I've been evaluating Freeplay as a potential solution for managing and evaluating our LLM-powered features. Our current volume is substantial, averaging well over 100k LLM calls per day across various models (GPT-4, Claude, and a mix of open-source via Bedrock). Most reviews and case studies I see focus on lower volumes or are from vendors themselves.

I'm specifically interested in hearing from teams operating at a similar scale. My primary concerns are:

* **Performance & Latency Overhead:** At 100k+ daily calls, even milliseconds of added latency for tracing and evaluation can become significant. Does the data collection and telemetry processing impact your p99 response times?
* **Cost of the Platform Itself:** While Freeplay aims to save money, its own pricing becomes a non-trivial line item at this call volume. Has anyone done a detailed TCO analysis comparing the platform cost to the potential savings from improved monitoring, prompt optimization, and reduced "bad" calls?
* **Pipeline Reliability:** With that volume of data flowing into their system for evaluation (especially if using custom evaluators or running batch evaluations), have you experienced pipeline backlogs, data loss, or significant delays in metrics availability?

Our initial testing at a smaller scale showed promise for consolidating our evaluation workflows, but scaling introduces new variables. I'm particularly keen on understanding:

* How you structured your integration—are you using the SDK, sidecar, or proxy?
* How you manage the storage and compute costs for the traces and evaluations, which are likely housed in your own cloud account.
* Whether the programmatic features (like test suites and automated evaluations) remain snappy and reliable with millions of traces.

Any concrete experiences, gotchas, or architectural insights would be invaluable.