Two Ukrainian developers, Marko Holyak and Lyubomyr Pytulya, launched the AI startup Talkruit, which brings together the entire IT hiring funnel in one window — from syncing responses from job boards to conversations with a voice agent and launching tests in a sandbox.

Startup co-founder Lyubomyr Pytulya told dev.ua how a pet project grew into a commercial product, how much development costs, how monetization works, and what the competition in the HR technology market is like.

How an ordinary work project grew into an AI startup for recruiting

The idea to create Talkruit was born during an ordinary work routine. Fullstack developers Marko Holyak and Lyubomyr Pytulya were working at the same IT company on a joint project, where they were creating a voice AI agent for conversations on various topics. Later, a practical question arose: what if such an agent were taught to communicate with candidates and prepare brief summaries about them in order to relieve the workload of recruiters?

When the first voice agent started working, the next problem arose — where to take the candidate next. Instead of the usual Google Meet or Zoom, the founders decided to create their own platform with flexible settings and manual control. Thus, HR and technical interviews appeared in the system.

An important catalyst was the company where Marko and Lyubomyr worked. It became interested in the development and began using it in real hiring — with all the imperfect scenarios and live data. It was there that the product was polished through trial and error, and later they also added automatic checking of test assignments so that developers wouldn't have to deploy a candidate's repository every time and run everything manually.

"Talkruit takes on the routine of hiring, from responding on Djinni to the interview report, while the decision about the candidate remains with the person," explains one of the startup's founders, Lyubomyr Pytulya.

When the system stabilized, the creators began doing cold outreach and offering free access. This is how they reached an outstaffing company that was developing a similar tool for its own needs. After the demo, the product seriously interested the team — and it became the first external client.

A team without a sales department and a launch on real vacancies

The startup has two co-founders — Lyubomyr Pytulya and Marko Holyak. There is no clear division of roles: tasks are distributed situationally, since there are always more ideas than time. They develop independently with the involvement of AI tools, and correspondence with clients, calls, and LinkedIn posts are also handled by the two of them — there is no separate sales or marketing department in the team.

They started writing code in mid-June 2026 in their free time from their main job. Already in the summer, the first version was working on real vacancies at the company where the founders worked, and later the first external client appeared. Currently, the entire hiring funnel goes through Talkruit: responses from Djinni, Work.ua, and Jobico are pulled in, resumes are analyzed, initial conversations with AI, HR and technical interviews, live coding are conducted, tests are checked, and final reports are generated.

Technology stack and what exactly AI does

The product backend is written in Node.js and TypeScript and runs on a server in Google Cloud, the cabinet interface is implemented on Next.js and hosted on Vercel. Data is stored in Postgres on Supabase with isolation for each company at the database level, and video communication works through LiveKit deployed on its own server.

The voice agent is built as a bundle of three services: Soniox is responsible for speech recognition, Gemini Flash generates responses, and Cartesia voices them. Each interview recording gets a separate virtual machine that is turned off after the call ends, and test tasks are collected and run in an isolated Cloud Run environment.

  • compares the resume with the job requirements;
  • conducts the first conversation with the candidate in Ukrainian or English (including checking English in the middle of a Ukrainian HR conversation) and forms a first impression;
  • analyzes directly the video of HR and technical interviews (not just a retelling) and sets timecodes;
  • assembles a project from a test assignment, writes and runs tests, checks the interface, and searches the code for access keys or passwords;
  • compiles the candidate's data into a single report.
  • The model returns the result according to a strict schema — if the structure is violated, the response is rejected. The technical interview score is reviewed or adjusted by the team lead, test scores are calculated by code according to fixed rules, and the final decision regarding the candidate always remains with a human.

    Opportunities in the hiring funnel

    The Talkruit platform unites the stages from receiving feedback to comparing finalists.

  • Candidate sources: automatic synchronization with Djinni, Work.ua and two-way integration with Jobico (including job posting and status transfer), import from Google Sheets, own job page, talent pool with auto-filled profile and a separate account for sourcers. The system prevents a single candidate from being simultaneously attached to multiple company vacancies.
  • Working with vacancies and CVs: AI helps formulate requirements and evaluation criteria, parses resumes, assigns tags (for example, "Courses"), and can automatically invite those who have passed the specified thresholds to a first conversation.
  • The first conversation with AI: the candidate communicates by voice in the browser at a convenient time using a template for a specific role. A separate AI controller checks the agent itself — whether it repeated questions, ignored the candidate's questions, or made up facts.
  • Interviews and live coding: independent selection and rescheduling of slots with synchronization to Google Calendar or Outlook, recording meetings in Talkruit, via a Google Meet extension or by uploading a recording from Zoom or Teams. In the built-in live coding editor (JS, TS, React, Python, Go, Java, C++), the candidate and team lead work together, and the level of permissions for the AI assistant for the candidate is configurable — from access to documentation only to full-fledged assistant mode.
  • Risk signal detector: during a call, the system records with time codes the moments that the interviewer should check (tab switching, pasting large code fragments, extraneous sounds or people in the frame, a sharp change in answer style), but does not automatically reject the candidate.
  • Competitors, money, and development plans

    The market already has voice AI recruiters (notably the Ukrainian project careeo.tech), classic HR systems such as Hurma or PeopleForce, where the first conversation is conducted by a human, as well as global solutions such as HireVue.

    “Our main difference is that the entire funnel is assembled in one place — from candidate feedback to the interview report, so the recruiter does not have to stitch together five different services. We work on a pay-per-actual-use model: no subscriptions or per-user fees — a company simply buys credits and spends them on specific operations, and we give new clients trial credits for 14 days. We also actually run test assignments in a sandbox, rather than just letting a language model read the code, we can record live interviews, including in Google Meet, and generate a report with timecodes. And most importantly — we first tested every function on real vacancies and refined it based on feedback from recruiters and team leads,” notes Pytulia.

    At the start, $300 in Google Cloud credits were enough — the developers tested the product in their free time without their own investments, and after the first client appeared, they received another $2,000 in credits through the Google for Startups program. Currently, the team is negotiating investments with one of the outstaffing companies.

    The monetization model is built on credits: $100 corresponds to 800 credits. For example, resume analysis costs about 7.5 cents, a minute of conversation with the AI — 12.5 cents, a minute of recording and analyzing an interview — about 6 cents, and checking a test assignment — $2.5. An outstaffing company of 50–200 people spends approximately $250–400 per month in the system.

    The founders' immediate plans are to connect card payments, expand integrations with ATS (Hurma, PeopleForce, later Greenhouse, Lever, Workable) and open an API, move to their own server for a larger number of simultaneous conversations, and prepare a version for the EU with data hosting in Frankfurt and compliance with AI Act requirements. In the spring, the startup plans to enter with its first clients in the US, and later in Poland and Eastern European countries with the addition of new languages.