_{Concept illustration. Actual project material appears below.}

IntiqAI

Multimodal hiring integrity. Evidence a reviewer can inspect.

A hiring platform spanning CV screening, skills assessment, AI assisted interviews, and HR reporting. Its integrity workstream turns session observations into timestamped evidence for human review.

My contribution: I owned the computer vision and multimodal integrity workstream, including identity continuity, session integrity, media analysis, and the evidence layer for reviewers.

Context: University of Prince Mugrin senior capstone, developed collaboratively by our team. Further development continues under university review.

5 integrity modules

2 modalities

Human review all hiring decisions

1st Place Best Prototype

Recognition

Best Prototype, 1st Place in the Computer Science track at the Makeen Annual Forum 2026, among senior capstone projects from the University of Prince Mugrin, Taibah University, and the Islamic University of Madinah.

What the integrity system reports

Each module reports independently and exposes its health status. Temporal confirmation helps prevent an isolated frame from becoming an event.

Module	Modality	Observation
Identity continuity	Visual	Changes from the identity enrolled at the start of the session
Presence anomalies	Visual	No face, multiple faces, and sustained person substitution
Liveness and spoof screening	Visual	Indicators of a live person or a presentation attack
Synthetic media screening	Visual	Deepfake and synthetic face indicators across multiple frames
Additional speaker detection	Audio	Indicators of another speaker in the session audio

From recording to review

flowchart LR
    A[Session recording] --> B[Normalization]
    B --> C[Integrity modules]
    C --> D[Timestamped events and evidence frames]
    D --> E[HR and Integrity Report]
    E --> F([Human reviewer decides])
    style F fill:#1e293b,stroke:#38bdf8,color:#e2e8f0

Loading

The HR and Integrity Report brings together the event timeline, captured evidence frames, module health, applied criteria, and plain language explanations. The design target is an HR reviewer who can inspect a flag without interpreting raw model output.

Integrity observations do not establish a person's intent, honesty, or emotional state. The platform does not automatically reject candidates. A qualified reviewer considers the evidence and makes the decision.

Reliability and evaluation boundary

The documented work includes five adversarial interview scenarios covering identity substitution, presentation attacks, and multiple people present.

Recording normalization addresses inconsistent browser frame rates so that temporal signals use elapsed session time correctly. Event persistence and visible module health are part of the reliability design.

These scenario checks describe the scope of testing reported here. Internal evaluation results and thresholds are not public, so this case study does not provide a public accuracy benchmark.

Technology

Area	Tools
Vision and identity	Python, OpenCV, InsightFace with SCRFD and ArcFace, MiniFASNet, Vision Transformers, MediaPipe
Audio and media	WavLM, FFmpeg
Platform	FastAPI, PostgreSQL, AWS S3

Public scope

This repository is a public project introduction. It documents the system, my contribution, the review workflow, and its boundaries.

Source code, models, thresholds, internal evaluation results, and candidate data remain private. Interview recordings and personal information are not published. No open source license is granted.

_{Mohammed Yousef Rasheed · GitHub · LinkedIn}