Candidate & Job Profile
Upload the Job Description and Candidate CV to calibrate technical questions
Job Description (JD)
Drop PDF, DOCX, Markdown, or YAML file here
Candidate CV / Resume
Drop candidate resume (PDF, DOCX, MD, TXT)
Question Plan & Strategy
Review auto-inferred seniority and select interview language to generate the question roadmap
Seniority Level
Auto-Inferred
Auto-inferred from JD / CV
Level will be automatically calibrated from uploaded documents
Interview Language
Option 1 of 2
Questions, rubrics, and voice synthesis will adapt to this language
Live Interview Room
Real-time technical interview with voice turn-taking and natural interruption
Ready
Connecting to interview session...
Ready
Gemini 3.8 Live Interview
Speech-to-Speech NativeReal-time bidirectional WebSocket streaming directly to Google Gemini Live API with sub-300ms latency
Evaluation & Comprehensive Report
Multi-dimensional rubric scoring, competency highlights, and hiring recommendations
Engine Parameter Tuning
Real-time adjustments for VAD sensitivity, barge-in thresholds, STT beam size, and dialogue flow
Parameter Impact Matrix & Tuning Reference
| Parameter | Increase ↑ | Decrease ↓ | Impact Area |
|---|---|---|---|
VAD_THRESHOLD |
Less sensitive — candidate must speak louder or clearer to be detected | More sensitive — picks up quiet speech, but higher risk of ambient noise | Microphone Sensitivity |
VAD_MIN_SPEECH_MS |
Must speak longer before speech is registered → fewer false triggers | Faster speech detection, but clicks or short noises may trigger speech | Speech Onset Detection |
VAD_HANGOVER_MS |
Allows longer pauses between words without cutting off → more complete answers | Faster response after candidate finishes, but risks cutting mid-sentence | Utterance End Detection |
VAD_BARGE_IN_MS |
Harder to interrupt agent — must speak longer before agent stops | Easier to interrupt — faster response, but speaker echo may cause false barge-in | Interruption Sensitivity |
VAD_BARGE_IN_THRESHOLD |
Requires clearer, louder speech to interrupt → resists speaker echo | Easier to interrupt even with quieter speech | Barge-in Threshold |
VAD_BARGE_IN_GRACE_MS |
Ignores microphone input longer right after agent starts speaking | Allows candidate to interrupt sooner after agent starts speaking | Echo Suppression Window |
STT_BEAM_SIZE |
Higher transcription accuracy but significantly higher latency (2x–5x) | Faster transcription response (greedy search) with minimal latency | STT Accuracy vs Latency |
TTS_SPEED |
Agent speaks faster → saves interview time | Agent speaks slower → clearer comprehension for listeners | Agent Speech Rate |
TTS_GAIN_DB |
Louder voice output — +6 dB ≈ doubles volume (too high may cause distortion) | Quieter voice output — harder to hear in noisy environments | Agent Volume Boost |
UTTERANCE_COLLECT_S |
Waits longer for candidate to continue speaking → full answers, slower flow | Faster response from agent, but risks splitting multi-sentence answers | Utterance Bundling |
NO_ANSWER_REMINDER_S |
More patient — waits longer during silence before prompting candidate | Prompts candidate sooner when silence is detected | Silence Reminder Delay |
MAX_STAYS_PER_QUESTION |
Agent stays on same question longer with multiple follow-ups → deeper inquiry | Agent advances questions more quickly → fewer repetitions | Question Depth Limit |
MAX_FOLLOWUPS_PER_INTERVIEW |
More follow-up probes allowed across the interview → dynamic conversation | Fewer probes → strictly follows predetermined question plan | Total Probing Allowance |