Gemini suggests, HR decide: an AI co-worker for fair and transparent hiring

One vacancy can bring hundreds of CVs.

In most Indonesian companies we work with, hiring still happens by hand: a job title that says little about the job, a post on a few channels, CVs read late at night, interviews remembered rather than scored, and an offer that nobody can explain a month later. Candidates wait for weeks without an answer. Owners wonder whether they picked the best person who applied, or only the one they remembered.

in HRKita we build HR and payroll systems for Indonesian companies at DataKita. For the Google Cloud AI Builder Cup we asked one question: what if every hiring decision had a co-worker who did the reading, kept the evidence, and explained its advice, while a person still made the call?

The rules that we set within the system

1. Gemini suggests; people decide. At every step the AI proposes and a named person chooses. Nothing moves a candidate forward or rejects them automatically.
2. Every score points to its evidence. A number without a quote from the CV, a test rubric line or an interview note is not shown.
3. The AI never sees what it must not judge. Name, photo, age, gender, religion, marital status and salary expectations are removed before any call.

One hire, step by step

An invented example: a company is hiring a Supervisor Akuntansi.

Knowledge, built together. The recruiter types a job title as vague as "Supervisor Akuntansi". Gemini suggests a full brief in seconds: the real activities (close the monthly books, reconcile cash, review journals), a 90-day target, and requirements in three tiers. The recruiter adjusts it: removes one requirement, adds a local tax task, rewrites the target. The screen shows what the AI proposed beside what the recruiter changed, and nothing goes further until the recruiter approves. From then on, the approved brief, not the model, drives the ad, the test and every score. The knowledge belongs to the organization.

Applying from a phone, then proving the skill. A candidate applies in about five minutes: a few knockout questions, their details, a CV, and a clear consent. Then a short technical test, such as an Excel case with a journal, etc. They can start now, or later on a laptop within 48 hours.

The screening. Gemini scores the CV against the approved brief, quoting the CV for every judgement, and scores the test answer against its rubric. A reviewer confirms. The candidate must pass both parts to continue.

Interviews as a team.  After the first interview, Gemini scores the candidate again and recommends whether to continue, with its reasons either way. The recruiter can follow the AI's recommendation. If HR disagrees with it, they write a short note explaining why. The hiring manager scores the second interview from a link on their phone, without seeing anyone else's score first. When all scores are in, HR sees the updated recommendation. The team discusses the candidate only after everyone has scored on their own.

The decision and the offer. Gemini suggests where in the salary band the offer should sit, based on the evidence; the app turns that into rupiah beside the budget, and the owner decides. Gemini drafts the offer letter and the rejection notices, and a person sends them.

Keeping it fair, and keeping it honest

We learned quickly that removing a name and a photo is not enough. People write things like "as a mother of two" inside their work history. So a person reads the redacted text before anything is sent to Gemini, and only the evidence sections of a CV are sent at all.

Candidates never see a score. They see their stage, their next step and its date, and they always get an answer.

What changed

On synthetic data, labelled as such:

• time from application to first answer: [before → after];
• recruiter minutes per CV: [before → after];
• decisions with a written reason: [before → after].

How it is built

One Cloud Run service; Gemini through the Google Gen AI SDK for the brief, the ad and posts, CV and test scoring, and the per-stage recommendations; storage for CVs and answers.  

What I learned

The hard part was not getting the model to score a CV, but deciding what it must never do, and building the checks that hold that line.

#AI hiring, Gemini, Google Cloud AI Builder Cup, Future of Work, recruitment, explainable AI, fair hiring, HR tech Indonesia