About this video
How we read PostHog data without opening PostHog. A chat interface built on the Javelin model, the AI SDK and Vercel Labs json-render, pulling live analytics on demand.
Full transcript
Edited transcript from the original Roboto page
So, Roboto has always been hell-bent on moving fast as a team. We do love PostHog, by the way, but we hate using it to read the data.
If you haven't heard about Javelin, which is the new kid on the block, it is one of the fastest models out there right now that you can use. It is cheap, fast, and just does the job. So, if you haven't checked Javelin, check it out. In this demo, what we're using behind the scenes is Javelin, the model. We're also using Vercel Labs JSON rendering, and we're using GPT Luma for the chat.
This is for our own use case. We've created this. It pulls data directly from PostHog, and all I'll have to do is click the button: "How many people visited the home page today?" Did you see how fast that was? Take a moment to appreciate that, first of all. The only bit where it took longer was the data coming from PostHog, but did you see how the decision was made from Javelin showing "Oh, it was 42 visitors." Simple as. There is no bells and whistles. The data was fast, the model is cheap, and it just does the job.
You can use it for a number of use cases. The main selling point here is it's cheap and it's fast and it just does the job. Also, worth mentioning that Vercel Labs JSON rendering is amazing for doing these kind of things. All the UI that you're seeing here is rendering its dynamic data, and you can see how fast it is, how easy it is to read, and you don't have to hire a data analyst. I mean, do if you want to, but just saying.
Read timestamped source captions
0:04Hey, how's it going? So, in today's
0:04Hey, how's it going? So, in today's demo, I'm going to show you another cool
0:06demo, I'm going to show you another cool
0:06demo, I'm going to show you another cool little feature that my team built for
0:08little feature that my team built for
0:08little feature that my team built for Roboto Studio. And this is about
0:11Roboto Studio. And this is about
0:11Roboto Studio. And this is about analyzing data from PostHog and [music]
0:13analyzing data from PostHog and [music]
0:13analyzing data from PostHog and [music] doing it as quickly as we can.
0:15doing it as quickly as we can.
0:16doing it as quickly as we can. So, Roboto has always been hell-bent on
0:18So, Roboto has always been hell-bent on
0:18So, Roboto has always been hell-bent on moving fast as a team. We do love
0:20moving fast as a team. We do love
0:20moving fast as a team. We do love PostHog, by the way, but we hate using
0:23PostHog, by the way, but we hate using
0:23PostHog, by the way, but we hate using it to read the data. [music]
0:25it to read the data. [music]
0:25it to read the data. [music] So, that being said, if you haven't
0:27So, that being said, if you haven't
0:27So, that being said, if you haven't heard about Javelin, which is the new
0:30heard about Javelin, which is the new
0:30heard about Javelin, which is the new kid on the block,
0:32kid on the block,
0:32kid on the block, it is one of the fastest model out there
0:35it is one of the fastest model out there
0:35it is one of the fastest model out there right now that you can use. It is cheap,
0:38right now that you can use. It is cheap,
0:38right now that you can use. It is cheap, fast, and just does the job. So, if you
0:41fast, and just does the job. So, if you
0:41fast, and just does the job. So, if you haven't checked Javelin, check it out.
0:43haven't checked Javelin, check it out.
0:43haven't checked Javelin, check it out. In this demo, what we're using behind
0:46In this demo, what we're using behind
0:46In this demo, what we're using behind the scenes is Javelin, [music]
0:47the scenes is Javelin, [music]
0:48the scenes is Javelin, [music] the model. We're also using Vercel Labs
0:51the model. We're also using Vercel Labs
0:51the model. We're also using Vercel Labs JSON rendering,
0:52JSON rendering,
0:52JSON rendering, and we're using GPT Luma for the chat.
0:55and we're using GPT Luma for the chat.
0:55and we're using GPT Luma for the chat. So, that being said, let me show you how
0:57So, that being said, let me show you how
0:57So, that being said, let me show you how this works. So, this is for our own use
1:00this works. So, this is for our own use
1:00this works. So, this is for our own use case. We've created this this pulls data
1:02case. We've created this this pulls data
1:02case. We've created this this pulls data directly from PostHog, and all
1:05directly from PostHog, and all
1:05directly from PostHog, and all I'll have to do is
1:07I'll have to do is
1:07I'll have to do is click the button, "How many people
1:08click the button, "How many people
1:08click the button, "How many people visited the home page today?"
1:10visited the home page today?"
1:10visited the home page today?" So,
1:11So,
1:11So, did you see how fast that was? Take a
1:13did you see how fast that was? Take a
1:13did you see how fast that was? Take a moment to appreciate that, first of all.
1:16moment to appreciate that, first of all.
1:16moment to appreciate that, first of all. The only bit where it took longer
1:19The only bit where it took longer
1:19The only bit where it took longer was the data coming from PostHog, but
1:21was the data coming from PostHog, but
1:21was the data coming from PostHog, but did you see how the decision was made
1:25did you see how the decision was made
1:25did you see how the decision was made from Javelin showing, "Oh, it was 42
1:28from Javelin showing, "Oh, it was 42
1:28from Javelin showing, "Oh, it was 42 visitors." Simple as.
1:30visitors." Simple as.
1:31visitors." Simple as. There is no bells and whistles, the data
1:33There is no bells and whistles, the data
1:33There is no bells and whistles, the data was fast, the model is cheap, and it
1:35was fast, the model is cheap, and it
1:35was fast, the model is cheap, and it just does the job. So, you can use it
1:37just does the job. So, you can use it
1:37just does the job. So, you can use it for a number of use cases.
1:40for a number of use cases.
1:40for a number of use cases. The main selling point here is it's
1:42The main selling point here is it's
1:42The main selling point here is it's cheap, and it's fast, and it just does
1:44cheap, and it's fast, and it just does
1:44cheap, and it's fast, and it just does the job. Also, worth mentioning that
1:47the job. Also, worth mentioning that
1:47the job. Also, worth mentioning that Vercel Labs JSON rendering is amazing
1:50Vercel Labs JSON rendering is amazing
1:50Vercel Labs JSON rendering is amazing for doing these kind of things. All the
1:52for doing these kind of things. All the
1:52for doing these kind of things. All the UI that you're seeing here is rendering
1:55UI that you're seeing here is rendering
1:55UI that you're seeing here is rendering its dynamic data,
1:56its dynamic data,
1:56its dynamic data, and you can see how fast it is, how easy
1:59and you can see how fast it is, how easy
1:59and you can see how fast it is, how easy it is to read, and you don't have to
2:02it is to read, and you don't have to
2:02it is to read, and you don't have to hire a data analyst. I mean, do if you
2:04hire a data analyst. I mean, do if you
2:04hire a data analyst. I mean, do if you want to, but just saying. Anyway,
2:07want to, but just saying. Anyway,
2:07want to, but just saying. Anyway, hope this kind of made sense, and you
2:08hope this kind of made sense, and you
2:09hope this kind of made sense, and you learned something new, and let me know
2:10learned something new, and let me know
2:10learned something new, and let me know if you have any questions, and thank you
2:12if you have any questions, and thank you
2:12if you have any questions, and thank you again for watching. Bye.
Questions answered
Editorial answers from the original Roboto page.
What is Javelin and why does Roboto use it?
Javelin is one of the fastest AI models available right now, and it's cheap too. We use it because it does the job without unnecessary complexity. The speed and cost efficiency let us focus on what matters: getting accurate answers from our data quickly.
Read the original ↗How fast is the demo you showed with PostHog data?
When I asked how many people visited the home page today, Javelin answered almost instantly with the correct number, 42 visitors. The only delay was pulling the data from PostHog itself, not the model processing it. That's the kind of performance we're talking about.
Read the original ↗What other tools are you using besides Javelin?
We're using Vercel Labs JSON rendering for the UI and GPT Luma for the chat component. The combination of these tools lets us build fast, readable interfaces that pull live data from PostHog and display results instantly without any bloat.
Read the original ↗Do you need a data analyst to use this setup?
Not necessarily. The interface makes it easy enough for anyone to ask questions about the data and get clear answers back. Of course, you can hire a data analyst if you want to, but this setup eliminates that requirement for basic queries.
Read the original ↗Why does Vercel Labs JSON rendering matter for this?
Vercel Labs JSON rendering is amazing because it handles dynamic data rendering efficiently. You can see how fast the UI updates, how easy it is to read the results, and it scales well without requiring heavy engineering on the frontend.
Read the original ↗What makes Javelin different from other AI models?
Javelin is cheap and fast, and it just works. There's no unnecessary processing or overhead. It makes decisions clearly and gets you the answer you need without any bells and whistles, which is exactly what we need when we're moving fast as a team.
Read the original ↗Can you use this approach for other use cases beyond analytics?
Yes, the setup is flexible enough for various use cases where you need to query data and get fast, clear answers. The main selling point is the combination of speed and cost, so any scenario where those matter is a good fit for this approach.
Read the original ↗