Topics

Hello, I'm Otsuka, CTO at Liberogic. I feel that we've entered an era where in-house app development can be done on a whim. ClaudeCode was a big catalyst for this.

The evolution of LLM and changes in costs

In recent years, the number of types and accuracy of LLMs (large-scale language models) has increased, and their cost performance has also improved.

OpenAI recently announcedgpt-ossIt has already been incorporated into various LLM hosting services.Groq of gpt-oss-120bLooking at the price list, it's a very cheap price.

Even with complex prompts, the cost is negligible if you run them a few hundred times per month.

The trigger was "Inquiry Summary"

Our website has an inquiry form, and we have long been using a system that uses AI to summarize the contents and send them to Slack.

Make your email form more convenient! It's best to let AI help you with spam and sales emails.

However, I was using a simple prompt using gpt-4o, and wanted to improve accuracy by using gpt-oss.

"Wouldn't it be interesting if the staff who actually respond to inquiries could create their own prompts?"

That's what I thought, but to test the prompt, I needed to paste past inquiry data into a Node program that would call the LLM API, run it, and verify it. While this is easy for an engineer, it's a bit of a challenge for a director.

So I came up with the idea,Creating an internal "playground" for AI servicesThat is

I created a single web application.

The specifications are very simple.

• Prompt input field on the left
• Load past inquiry data stored in Notion
• Call the Groq API and display the results on the right

Although it is just this function, it creates an environment where directors can try it out right away.

In the past, it took a long time to prepare the infrastructure and build the servers even for an internal tool of this scale, but nowCloudflare WorkersYou can quickly set it up in an environment like this. Furthermore, once you have determined the program requirements, ClaudeCode will create it for you in just a few minutes.

It only took about two hours

It only took about two hours from coming up with the idea to completing it*.

It was released on the same day, and everyone in the company, including our director, was able to try it out right away.

What impresses me isn't that non-engineers can now use AI.

"Because we can quickly create in-house applications, we can involve not only engineers but also people around us in the development process." --This is the most interesting part.

You can create purpose-specific tools in a short amount of time and improve them as the whole team tries them out.

I feel that this kind of movement will make future projects even more exciting.

*The thing that took me the longest was not the programming part, but figuring out how to set up client-side scripts in Hono + Vite.