DeepScreenDeepScreen
Funnel

Coding & Skill Assessments

DeepScreen's assessment pipeline is gated — candidates must clear each stage before advancing. This ensures expensive recruiter time is reserved for candidates who have already demonstrated technical competence and communication skills through automated, proctored evaluation stages.

Assessment Pipeline

1

Coding Assessment

Judge0 CE + Tauri

Candidates solve timed coding challenges in a sandboxed Judge0 CE environment. Tauri desktop client provides OS-level process monitoring — detects AI copilots (ChatGPT desktop, GitHub Copilot), remote desktop tools (TeamViewer, AnyDesk), and browser automation (chromedriver, geckodriver). Code is compiled and tested against hidden test cases with time complexity analysis.

G

Gate: Score Threshold

Configurable

Coding score must meet the minimum threshold for the specific role. Thresholds are configurable per position — a senior backend role requires 80%+ while a junior frontend role may accept 60%+. Candidates who pass advance to Stage 2. Candidates who fail receive a detailed performance report and are removed from the pipeline.

2

AI Avatar Interview

sherpa-onnx + Supertonic 3

3D avatar conducts a structured interview with adaptive questioning powered by Groq LLM (350-600ms response). Speech handled entirely offline: Web Speech API for STT, Supertonic 3 ONNX INT8 for TTS (99M parameters, 31 languages, CPU-only). LLM adapts question difficulty based on response quality — strong answers lead to harder follow-ups, weak answers trigger clarification questions.

G

Gate: Interview Score

Threshold Check

AI-assessed interview score must meet threshold. The LLM evaluates technical depth, communication clarity, and problem-solving approach. Scores include confidence intervals rather than single-point estimates — a candidate with 0.75 ± 0.05 is more reliably scored than one with 0.75 ± 0.15.

3

Live Recruiter

LiveKit SFU

Multi-participant WebRTC session via LiveKit SFU. Audio latency: 80-150ms. Video latency: 120-250ms. Multi-user sync: <300ms. CPU usage: <15% with 4 participants. Recruiter sees the full assessment history and can ask targeted follow-up questions based on specific coding or interview performance gaps.

R

Final Report

Evidence Pack

Comprehensive report combining: resume analysis summary, coding assessment scores with test case breakdown, AI interview transcript with confidence scores, proctoring evidence pack (gaze logs, head pose, object detection, process monitoring), and live interview notes. Weighted composite score with configurable weights per role type.

Proctoring Signals

SignalMethodThreshold
Gaze Tracking
MediaPipe Face Mesh (468 landmarks)Off-screen >30% frame time
Head Pose
solvePnP with facial landmarksYaw/pitch ±20° for 5+ frames
Face Presence
MediaPipe detectionAbsent for 3+ consecutive frames
Multi-Person
YOLOv8s ONNX INT82+ faces detected for 2+ frames
Prohibited Objects
YOLOv8s ONNX INT8Phone/tablet confidence >0.5
Process Monitoring
Tauri sysinfo crateBlocklist match → immediate flag

Offline TTS Specifications

ParameterValue
Parameters
99 million
ONNX Model Size
404 MB total
Languages
31
Runtime
CPU-only, no GPU required
Integration
sherpa-onnx Rust bindings
Init Latency
<50ms
Privacy
Fully offline, no network dependency