AI/ML Engineer Assessment: Multimodal Submission Integrity & Plagiarism Detection Engine
AI/ML Engineer Assessment: Multimodal Submission Integrity & Plagiarism Detection Engine
Project Constraints and Rules
Permitted Tools and AI Disclosure Rules
Time Budget and Compute Limits
Multimodal Ingestion and Normalization
Text Extraction from PDF, DOCX, and MP4
PDF Text Extraction with Page Tracking
DOCX Paragraph and Table Extraction
Audio Track Extraction and Speech Transcription
Extracting Audio Tracks from MP4 Files
Speech Transcription with Whisper-tiny or Vosk
Whitespace Normalization and Header Footer Stripping
Whitespace and Unicode Normalization
Detecting and Stripping Repeated Headers and Footers
Token-Based Chunking with Overlap
Source Metadata Preservation Through Chunking
Handling Malformed, Zero-Byte, and Silent Files
Indexing and Internal Peer-to-Peer Similarity
Dense Sentence Embeddings
Sentence-Transformers with all-MiniLM-L6-v2
Vector Index Options
FAISS Flat Inner-Product Index
File-Backed ChromaDB Collections
NumPy Cosine Similarity Matrix
Cosine Similarity and Thresholding
Document-to-Corpus Matching
Per-Document Peer Similarity Percentage
Self-Match Exclusion
Multi-Signal Integrity Detection
AI-Content Likelihood Estimation
Transformer-Based AI-Text Detector Classifier
Perplexity and Burstiness Heuristics
Reproducible Mock Detector Wrapper
External Plagiarism Simulation
Local Reference Knowledge Base
Optional Free Search API with Offline Fallback
Shared Boilerplate and Prompt Header Discounting
Combining Signals into a Structured Score Breakdown
Citation and Explainability Engine
Exact Matching Passage Citation Pairs
Citation Metadata and Chunk Similarity Scores
Per-Document JSON and Markdown Reports
Machine-Readable JSON Report Schema
Human-Readable Markdown Report Layout
Penalty Justification Rule
Report Design for 30-Second Cross-Verification
10,000 Document Scaling Brief
Compute Choices: CPU versus GPU
Asynchronous Processing Queues
Vector Database Selection and HNSW versus Flat Indexing
Cost, Throughput, and Latency Bottlenecks
Sample Submission Dataset
Corpus Composition Requirements
Planted Duplicate, Cross-Modal, LLM, and Original Cases
Degraded and Malformed Test Files
dataset_manifest.json Ground Truth Format
Test Scenarios
Scenario 1: Cross-Modal Duplication
Scenario 2: Paraphrased Content Detection
Scenario 3: Corrupted and Empty Submissions
Scenario 4: Benign Common Formatting
Scenario 5: Multi-Document Batch Evaluation
Repository Deliverables
Modular Source Layout and Runner Script
README Contents and Assumptions Section
EVALUATION.md Structure
Precision, Recall, and Confusion Matrix Reporting
Failure Analysis Write-Ups
Trade-off Analysis
Requirements, Model Scripts, and Configuration Files
Repository Submission and Deadline
Loading SophistAI