Anyone actually using Helicone in production with 10k+ daily requests?

Hey everyone, been lurking here for a bit and finally decided to post. I've been evaluating Helicone for our mid-sized deployment where we're serving AI features. We're currently averaging around 15k-20k requests per day across a mix of OpenAI and Anthropic models, and I was really drawn to Helicone's promise of unified logging and cost analytics.

Our main stack is on AWS with Terraform-managed EKS, so I was curious if anyone else is running at a similar scale in a production environment. I'm specifically wondering about a few things:

* **Performance & Latency:** Did you notice any meaningful added latency? We're sensitive to end-user response times.
* **Cost Tracking Granularity:** How well does the cost breakdown work with custom models or when you have significant caching layers (like Redis) in front? Our current homegrown dashboard is... lacking.
* **Infrastructure Overhead:** We deployed the CloudFormation stack for the AWS integration. Any pitfalls as request volume grew? I'm thinking about DynamoDB capacity or Lambda concurrency limits.

Here's a snippet of how we integrated it with our Python service, pretty straightforward:

```python
import openai
from helicone.openai_proxy import openai

openai.api_key = "your-openai-key"
openai.helicone_api_key = "your-helicone-key"
openai.base_url = "https://oai.hconeai.com/v1"

# Proxied request goes to Helicone
response = openai.ChatCompletion.create(
model="gpt-4",
messages=[{"role": "user", "content": "Hello"}]
)
```

The dashboard is slick for sure, but I'm more interested in the operational side. Have you hit any scaling walls? How's the support been if you ran into issues? Also, if you're using Terraform, did you end up managing any Helicone resources as code, or just use their provided templates?

Would love to hear real numbers or war stories before we fully commit. The idea of offloading our monitoring and cost allocation is tempting, but I need to be sure it's robust.

-- Amy