AI Technical Screening for Hiring

No. Talkruit doesn't replace the recruiter or the tech lead — it closes the gap between them. A recruiter can't filter out technical incompetence because they don't understand the details themselves. AI does the initial check against the company's criteria, and the tech lead verifies the conclusion. The decision is always made by a human.

The score is made up of several factors: how well the CV matches the actual answers, contradiction detection between what's written and what's said, the quality of answers to technical questions, and suspicious moments with timestamps. After AI screening, the tech lead personally verifies the technical part — only after their confirmation is the score considered final. AI structures the information, the human makes the decision.

After a LiveKit call with the candidate, Gemini analyzes the video recording fragment by fragment — without a separate transcription service — and assembles a report across tabs: skills 1-10 per technology, strengths and weaknesses, an overall conclusion. The tech lead opens the report, can change the score, add a comment — and only after confirming does the report go to HR. Instead of writing everything from scratch after a 60-minute conversation.

The candidate picks a convenient slot themselves from what's available in the tech lead's calendar — no calls or emails between the recruiter and tech lead. If something changes, the candidate reschedules it themselves, and all participants get an email and choose a new time. Notifications are sent at every step: confirmation, reminder, time change.

Speech recognition (Soniox) and voice synthesis (Cartesia) natively support Ukrainian, while Gemini Flash-Lite handles understanding answers and generating questions — not one universal model, but the best solution for each specific task. Detecting when a person has finished speaking works in two steps: a local model (Silero VAD) instantly catches the moment of silence, then a separate lightweight Gemini check confirms whether it's really the end of a thought and not a pause to think — so the agent doesn't interrupt mid-sentence like typical voice bots. The candidate speaks naturally, with dialect and technical terms — the model understands the context. English is also supported for international candidates.

Usually 3-7 minutes depending on the role and question depth. Configurable per job. The candidate opens the link, goes through a quick check (microphone, quiet room), and starts the conversation right away — no queues or registration.

Yes. The company configures the questions, evaluation criteria, and interview tone. You can set custom instructions for the AI — domain specifics, evaluation style, special requirements for the role. Each job can have its own configuration.

In Google Cloud Storage in European regions. Encrypted with standard AES-256 (Google-managed keys). Access is limited to your organization via Supabase Row-Level Security isolation.

You top up your balance with credits ($100 = 800 credits, the $500 package comes with a +10% bonus) and pay only for completed AI operations. Calls are billed per minute: AI screening — 1 credit/min (a typical 7-minute one = 7 credits, less than a dollar), a recorded tech or HR interview with AI summary — 0.5 credit/min (40 minutes = 20 credits). AI CV analysis — 0.6 credit, job generation — 1.5 credits, test task evaluation in a sandbox — 20 credits. No subscription or per-user fees: dashboard, analytics, configuration, calendar, Telegram chat with candidates, and unlimited users are free. For context: Western AI screeners charge $2.5–4 per screening; ours costs ~$0.88. At the start we give a 14-day trial with 150 credits (expiring after 14 days), and purchased credits never expire; no charge for missed calls or technical failures.

The candidate and the tech lead see the same code in real time — no manual screen sharing, tab switching, or “wait, let me share my screen.” Every code change is visible the moment the candidate makes it. A library of ready-made tasks for a specific stack and difficulty level means there's no need to improvise a task live or search for one mid-call — pick the stack and level, and the task is ready.