Velvet

Hey Product Hunt 👋 Emma and Chris here from Velvet. We built an AI gateway to warehouse OpenAI and Anthropic requests to your database. Engineers use Velvet logs to analyze usage and costs, evaluate models, and generate datasets.

Our first product was an AI SQL editor. It worked really well, but we had limited insights into what happened between our app and OpenAI. We started storing requests to our database, giving us the ability to query data directly. It gave us so much more control over developing our AI features.

We didn’t think of it as a product until one of our customers (Find AI) asked to use it. We warehoused over 3 million requests for them in the first week, logging up to 1,500 requests per second during their launch. Now we have many startups using Velvet daily for caching, evaluations, and analysis of opaque endpoints like the Batch API.

Try a sandbox demo → usevelvet.com/sandbox (no signup needed)

Read the docs → docs.usevelvet.com

Get started with two lines of code 🧑🏻‍💻

It's easy to get started! Once set up, we warehouse every request to your database.

Requests are formatted as JSON. You can include custom metadata in the header - like user ID, org ID, model ID, and version ID. This means you can run complex SQL queries unique to your app. Granularly calculate costs, evaluate new models, or identify datasets for fine-tuning.

We'd love to hear from you! Email team@usevelvet.com with any feedback or questions.

Watch a demo video to learn more.

Congratulations on launch! Our company Find AI has been an early customer of the AI gateway, and it's been a game-changer for our engineering team to interact with OpenAI. Here's some of the things we do with the millions of requests we warehouse weekly with Velvet: - Cost analysis: What's AI spend per user, per search, for logged out users, and in prod vs. development? - Debugging: When a query does something weird, trace it back to the OpenAI call. We traced one query that occasionally returned gibberish, and realized that the temperature was just too high. - Caching: particularly use this on smoke tests, so that we can actually query OpenAI if our prompt/model/etc changes to verify expected behavior, but skip paying money if things are the same - Testing model upgrades: This week we've been exporting data of some models in gpt-4o, then re-running the exact queries against gpt-4o-mini. To our surprise, a couple of our classification queries had 100% the same accuracy on the mini model, so we could immediately switch and save money.

Congrats for the launch and thank you, we are happy users of Velvet, it's not only very simple to set up and use but also highly reliable and very well designed. Kudos team 🙌