A neural network for HR: job posts, candidate emails and resume review

A neural network for HR saves time on text routine: job posts, emails, interview questions, first-pass review of a stream of resumes. But this field has two traits absent in marketing or support. First: resumes and forms contain personal data of real people. Second: decisions about people depend on the result, so a model's bias turns into real harm. Below: what to hand a model, how to treat data, how to build resume review so the decision stays with a human, and what it costs.

A neural network for HR: which tasks suit a model

A guide: the fewer the consequences of a mistake and the easier it is to verify the result, the better the task fits.

For job posts and emails you do not have to take the most expensive model: start with an inexpensive one and test it on your own examples. How to choose a model by price and strength index is in the article on the model strength index↗. There is also a separate piece on texts in general: a neural network for writing texts↗.

Job posts and emails: what to ask for

A good request contains facts, not a plea to "write it nicely": the position, duties, required and desirable requirements, terms, the salary range (if you publish it), work format. Add constraints: "do not mention age, gender or marital status, do not use words like 'young' or 'energetic'". And ask to separate the required from the desirable: it is easier to match resumes against that.

For candidate emails set the tone and structure. A rejection email: thanks, a clear decision without needless detail, no promises like "we will keep your resume" if you will not keep it. A model will prepare a draft of such an email, but send it yourself: automatic rejection mailings sent unread quickly lead to awkward situations.

Candidates' personal data

A resume is a full name, contacts, date of birth, workplaces, sometimes a photo, marital status and an address. Everything you pass to a model in a request goes to the provider's side. Practical rules:

1. Pass the minimum. To match against criteria you do not need the name, phone, email and address. Remove them before sending and label the candidate with a number. How to replace data with labels and restore them, we covered in the article on logging without leaking data↗.
2. Do not send extras. A photo, health data, marital status and nationality are not needed to assess skills, and even less in a request.
3. Clarify the legal side. Processing candidates' personal data is governed by the personal data law: consents, purposes, storage, transfer to third parties. We do not give legal opinions; discuss questions about passing data to a model API and about where it is stored with the company lawyer before launch.
4. A separate key. Create a key for HR: in the dashboard you set a daily and monthly limit on it and a list of allowed models (more about key limits↗). That way HR spend is visible separately and stays under its ceiling.
5. Store results carefully. A model's answer about a specific candidate is also personal data and should live where the resume lives.

Bias: how not to hand the decision to a model

A model is trained on texts that carry human stereotypes. So resume review must be built to leave it no room for a "gut feeling" score.

• Criteria are set in advance and taken from the job post. Do not ask "rate the candidate on a ten-point scale". Ask it to mark each criterion as met, not met or unclear and to give a quote from the resume.
• No protected attributes. Remove the name, gender, age, photo, marital status and nationality before sending to the model.
• The model does not reject. Its output is a summary for the recruiter, not a decision. A human rejects.
• A swap test. Take one resume and send two versions with different names or a different gender. The results per criterion must match. If they do not, you have a problem and such a process cannot be used. Repeat the test whenever you change the model or the prompt.
• Spot audits. Once a month a recruiter manually reviews random resumes that received "not met" and looks for strong candidates who were missed.

Example: a resume summary against criteria

import json
import os

from openai import OpenAI

client = OpenAI(
    base_url="https://api.zukko.pro/v1",
    api_key=os.environ["ZUKKO_API_KEY"],
)

CRITERIA = [
    "At least 3 years of experience with PostgreSQL",
    "Experience writing REST APIs in Python",
    "Experience leading a project in a team of 3 or more",

The quote_found field catches the nastiest failure: an invented quote. If the model cites a phrase that is not in the resume, the result for that criterion cannot be used. The unclear status does not mean "a weak candidate"; it is a signal for the recruiter to look themselves.

What it costs

Prices on October 18, 2026, per 1M tokens. Suppose a stream of 200 resumes. A resume and form take about 1,000 tokens, criteria and instruction about 500, the answer about 250. Input is 200 × 1,500 = 300,000, output 200 × 250 = 50,000.

• gpt-6.1-sol (input $0.27, output $1.32): 300,000 × 0.27 / 1,000,000 + 50,000 × 1.32 / 1,000,000 = 0.081 + 0.066 = $0.147.
• claude-sonnet-5-5 ($0.4 / $2): 0.12 + 0.10 = $0.22.
• deepseek-v4.1-flash ($0.013 / $0.038): 0.0039 + 0.0019 ≈ $0.006.

That is the cost of a run over 200 resumes. The model price is secondary here: what matters is that the process is verifiable. If you rerun the same resumes with the same criteria, turn on the response cache with the X-Zukko-Cache: true header: an identical request within 7 days is served from the cache and not charged. Current prices: GET https://api.zukko.pro/v1/models.

FAQ

Can a neural network be trusted to select candidates?

Not as the final decision. A model is good for a summary against set criteria, and the decision must be made by a person who can explain it to the candidate.

Which model for HR tasks?

For emails and job posts you can start with claude-sonnet-5-5 or gpt-6.1-sol and check the result on your own texts. For first-pass review compare two or three models on 20 resumes and check the stability of the result, including with the name-swap test.

How do I avoid sending personal data to a model?

Remove names, contacts and addresses before sending and label the candidate with a number. If former employers' names and job titles remain in the text, that is context that may identify a person too: decide whether the task needs it.

How do I connect from Russia?

Access from Russia goes through the gateway, payment by SBP, a MasterCard card or crypto. The key is created in the Zukko dashboard↗.