DeepScreenDeepScreen

DeepScreen

Intelligent Hiring Platform

Revolutionary hiring intelligence platform powered by Rust-native inference and AI. Screen candidates 3-5x faster with semantic matching and real-time assessment.

Powered By

Next.js
Django
Python
Rust
ONNX
Qdrant
PostgreSQL
Embedding
Hugging Face
Rocket
Tauri
Next.js
Django
Python
Rust
ONNX
Qdrant
PostgreSQL
Embedding
Hugging Face
Rocket
Tauri
Next.js
Django
Python
Rust
ONNX
Qdrant
PostgreSQL
Embedding
Hugging Face
Rocket
Tauri
DeepScreenmatchestoptalentwithprecisionAI.OurenginerunssemanticCVparsingagainstsection-weightedembeddings,adaptivecodingchallengeswithreal-timeplagiarismdetection,andliveintegrity-verifiedinterviewspoweredbyon-deviceASRandTTS—allunifiedinasingleplatformthatdeliversfair,fast,anddata-drivenhiringdecisionsatscale.
DeepScreenmatchestoptalentwithprecisionAI.OurenginerunssemanticCVparsingagainstsection-weightedembeddings,adaptivecodingchallengeswithreal-timeplagiarismdetection,andliveintegrity-verifiedinterviewspoweredbyon-deviceASRandTTS—allunifiedinasingleplatformthatdeliversfair,fast,anddata-drivenhiringdecisionsatscale.
50K+

CVs Parsed

99.7%

Platform Uptime

8K+

Coding Assessments

<50ms

Vector Search

2.5K+

AI Interviews

150ms

Audio Latency

End-to-End Platform

Everything in One Place

From CV matching to final hire — a unified pipeline powered by Rust-native inference

Philosophy

Smarter screening.
Fairer hiring.

DeepScreen doesn't just match keywords — it understands context. Every candidate is evaluated through semantic reasoning, adaptive challenges, and integrity-aware interviews. Built for teams that refuse to compromise on talent quality.

Rust • Tauri
ONNX INT8
Qdrant gRPC
Python
LiveKit SFU
MediaPipe
01

Semantic CV Matching

Section-wise embeddings across Skills, Experience, Projects, and Summary using all-mpnet-base-v2 ONNX INT8. Four Qdrant gRPC searches per JD with weighted aggregation — 0.35 Work Experience, 0.30 Skills, 0.20 Projects, 0.15 Summary. Precision@10 improves by 11.4pp over whole-doc retrieval, ensuring the strongest candidates surface first.

02

Coding Assessment

Adaptive coding challenges powered by Rust-native execution sandbox. Supports Python, JavaScript, C++, and Rust with real-time test validation. Questions dynamically adjust difficulty based on candidate performance. Integrated plagiarism detection analyzes solution structure, not just output — catching AI-generated and copy-pasted submissions.

03

AI Interview System

Offline-first STT/TTS pipeline using sherpa-onnx (Zipformer ASR + VITS TTS) runs entirely on-device via Tauri. Real-time 3D avatar with lip-sync delivers adaptive questioning that probes skill gaps identified during CV matching. Questions evolve mid-interview based on candidate responses, creating a truly dynamic assessment.

04

Live Video Interview

LiveKit-powered SFU architecture delivers 80-150ms audio latency and 120-250ms video latency with true multi-user synchronization. Built-in recording, screen sharing, and real-time annotation tools. Interviewers see live AI-generated suggested questions and candidate competency metrics alongside the video feed.

05

Proctoring & Integrity

Multi-layer integrity monitoring: MediaPipe gaze tracking detects off-screen focus, YOLOv8s INT8 identifies unauthorized objects and additional persons, OS-level process monitoring (Tauri Rust) flags forbidden applications. All detection runs locally — no video stream leaves the device. Configurable sensitivity per assessment tier.

06

Cheating Prevention

Advanced behavioral analysis correlates gaze patterns, keystroke dynamics, and response timing across all assessment stages. Cross-references candidate identity via continuous facial verification. Generates an integrity score for each candidate that feeds into the final ranking — making cheating detectable even when individual flags are subtle.

How Shortlisting Works

Three-Stage Ranking Pipeline

Retrieve → Re-rank → Score — cheapest operation always runs first

01

Job Description

Parse & Embed

The job description is split into four structured fields — Work Experience, Skills, Projects, and Summary — each embedded into a 768d vector using all-mpnet-base-v2 ONNX INT8 (35MB). Work Experience receives the highest weight (0.35) as the strongest predictor of job performance, followed by Skills (0.30), Projects (0.20), and Summary (0.15). Missing sections default to 0.0. This section-wise approach improves Precision@10 by 11.4pp over whole-doc embedding by preserving field-level semantic resolution.

02

HNSW Vector Retrieval

Stage 1: Broad Semantic Search

Four Qdrant gRPC searches are issued in parallel, one per embedded field, each with a has_id() filter scoped to that collection's applicant vectors. Results are aggregated using the same section weights from Stage 1. HNSW settings: COSINE distance, m=16 neighbours, ef_construct=200, ef=128 — delivering Recall@10 of 0.93. The top 2K candidates by weighted vector score move to Stage 2. Because vector search is O(log n) and runs first, the expensive stages operate on a dramatically reduced pool.

03

Cross-Encoder Re-ranking

Stage 2: Deep Transformer

For each of the 2K candidates, four (job-field, resume-field) pairs are tokenized, padded to max_len, and stacked into a single [4 × max_len] input tensor. A single ONNX Runtime session.run() processes all four pairs at once via intra-op parallelism — approximately 3x faster than four sequential calls. The model is ms-marco-MiniLM-L12-v2 quantized to INT8: 134MB → 34MB (74.6% reduction) with negligible accuracy loss. Four output logits are weighted identically to Stage 1 weights.

04

BM25 Lexical Re-ranking

Stage 3: Exact Keyword Recovery

Cross-encoder scoring operates on LLM-structured metadata that may omit specific tokens such as version numbers (Python 3.12), certification codes (AWS-SAA-C03), and proprietary tool names. BM25 on raw_full_text recovers these exact matches. Parameters: k₁ = 1.2 (term frequency saturation), b = 0.75 (length normalization). Tokens are lowercased and split on non-alphanumeric boundaries with minimum token length 2. This lexical signal captures what semantic embeddings inherently smooth over.

05

Score Blend

Weighted Fusion

All three scores are min-max normalised within the top-K pool (neutral 0.5 when all candidates are tied). The final score is computed as: score(u) = 0.65 · c_n(u) + 0.20 · v_n(u) + 0.15 · b_n(u). Cross-encoder receives the highest weight (0.65) because it jointly processes both texts through the full transformer — capturing deep semantic interactions that bi-encoder and lexical methods miss. The blended score drives the final ranking.

06

Ranked Shortlist

Final Output

Candidates are sorted by blended score in descending order. Each entry carries its per-stage scores, competency breakdown, and integrity verification flags for downstream processing. Internal benchmarks on 500 candidates across 50 job descriptions show: Precision@10 of 0.76 (vs 0.68 whole-doc), Recall@10 of 0.83 (vs 0.74), and NDCG@10 of 0.79 (vs 0.71). The shortlist feeds directly into the AI Interview pipeline for adaptive questioning based on identified skill gaps.

Platform Capabilities

Powerful Features

Purpose-built components designed for modern hiring workflows

Core Engine

Rust-Native Inference

In-process ONNX models eliminate Python GIL overhead. 3-5x faster inference with 60-80% memory reduction.

Learn more
Intelligence

Semantic Resume Matching

Section-wise embeddings across Skills, Experience, Projects, and Summary. Qdrant gRPC with 11.4pp precision gain.

Learn more
Assessment

AI Interview System

Offline STT/TTS with sherpa-onnx. Real-time 3D avatar with adaptive questioning based on skill gaps.

Learn more
Integrity

Advanced Proctoring

OS-level process monitoring, MediaPipe gaze tracking, YOLOv8s object detection. Tauri-powered integrity.

Learn more
Accuracy

Multi-Stage Ranking

Three-stage pipeline: HNSW vector retrieval → Cross-encoder re-ranking → BM25 lexical scoring.

Learn more
Communication

LiveKit Integration

SFU-based real-time video interviews. 80-150ms audio latency, 120-250ms video, true multi-user sync.

Learn more
Performance

Edge-Optimized Inference

Supertonic compression via sherpa-onnx — all models run offline on consumer hardware

69MB

Total Model Size

all 6 models combined after INT8 quantization

4.2×

Compression Ratio

supertonic compression via sherpa-onnx edge optimization

270MB → 69MB

Memory Reduction

74.4% total memory saved across ASR, TTS, and encoder models

RTF 0.05

Zipformer ASR

real-time factor on device — 20× faster than real-time speech

<150ms

End-to-End Latency

STT + TTS pipeline on consumer hardware (M1 / i7)

35MB

all-mpnet-base-v2

optimized INT8 ONNX — fits in L2 cache for sub-ms inference

Ready to transform your hiring?

Join leading organizations using DeepScreen to find better talent faster. Start your free demo today.

0

Active Users

0

Candidates Screened

0.0%

Platform Uptime

0

Support

DocumentationGitHubContact