CariKerjaIT Job AI

testing on google
testing on line
job desc preview
location preview

Inspiration

The project inspiration is coming from the recent changes on the tech industry in Indonesia. The Indonesian tech bubble has passed and for the past 1 year or so there are lots of lay offs from companies of all sizes. Lots of startup closing down business. Then come up this idea to help user find a suitable jobs that is best for them, whenever and wherever they like to search for it. Our product provides location, work type, and salary insights for many available jobs in Indonesia.

What it does

It is a bot that you can install by scanning QR Code and add it to your Line Messaging app. Line messaging app is the most popular messaging app for younger generation in Indonesia. Line has 30 million users in Indonesia (2nd in the market)[1] which makes it a good market for Job portal bot.

Our vision is to make natural language as first-class interface to find jobs and help users with slow internet in Indonesia get the same access like the one that does. So no need for high bandwidth internet to get the access and open it in your desktop/mobile. Just text it.

Main functionality

The main functionality of the bot should be to guide user to find jobs in CariKerjaIT then also help user if they need some training before applying or while applying so they can be ready in the job. We use Class Central to find the best courses online as well as local player called Skill Academy.

How to try

Using Line:
- You can visit this page Add to line
Using website: https://ai.carikerja.it
- Demo: https://www.loom.com/share/f60f288eab9f4f9c84e31189633032fc

How we built it

We have the bot 100% built using the Vertex AI Agent builder where we add a couple of tools including CariKerjaIT API, Class Central API and SkillAcademy API. The CariKerjaIT itself is a web portal written in Ruby on Rails, Elm and some geocoding using Google Maps API.

The bot now has 2 interfaces, one is Line for our end user and Slack for our internal testing.

Challenges we ran into

We encounter an issue where sometimes the bot just refuse to reply to us which one of the issue is the token limit. Although in some corner cases the bot from Vertex still refuse to reply even when we are still on the token limit.

In the first one week during the hackathon we would like to use Vertex AI Search but run into cost issue, where our Vector Database is just too expensive for us to run. So we need to rethink our approach.

To limit our usage and control it we also introduce budget alerting in Google Cloud so that we are not run into cost problem.

Accomplishments that we're proud of

Integrate our CariKerjaAPI with the bot agent in Vertex AI, integrates it with Line and Slack as well as having the happy path working. We also version control our bot in Git so we can work in parallel and avoid issues.

What we learned

Please be mindful when using AI API/products online, it can be costly. So plan and budget accordingly. Install alerting is a must.
Make many test cases so it would be easy for us to test.
The biggest insights that we have is the AI can be very powerful when we give them lots of tools. And we can package them to create a marketable product.

What we hope it is available in Vertex AI Agent

Options to seed multiple test cases and rate the test results.

What's next for CariKerjaIT Job AI

Test our app with real users, we would like to open some beta access to selective customers and get feedback how to improve it a bit more.
Improve the bot to add more edge cases like when conversation becomes open ended.
Wrapping the Vertex AI Agent with our custom made wrapper so that we can handle some edge cases ourself and not 100% relying on native integration that Vertex AI Agent provides.