AI/ML Engineer Assessment: Multimodal Submission Integrity & Plagiarism Detection Engine

This assessment asks you to build a pipeline that checks submitted documents and recordings for copying, reuse, and machine-written text. The system must read PDF, DOCX, and MP4 files, compare them with one another and with a local reference corpus, and explain every finding with a location a reviewer can verify. It matters because integrity decisions carry real consequences, so a flag is only useful if it can be defended with evidence.

Project Constraints and Rules

This branch sets the 72-hour time budget, the CPU or free-tier GPU compute limits, and the rules for disclosing AI coding assistants and avoiding paid detection or plagiarism APIs. Its child articles detail each rule. Every later branch assumes these limits hold.

Multimodal Ingestion and Normalization

This branch covers converting PDF, DOCX, and MP4 inputs into clean, chunked text that keeps file, page, and timestamp metadata. Its children cover per-format extraction, normalization, chunking, metadata preservation, and handling of malformed or silent files. Every downstream stage depends on the quality of this step.

Indexing and Internal Peer-to-Peer Similarity

This branch covers embedding chunks, storing them in a vector index, and measuring how much each submission overlaps with the others in its batch. Its children cover embedding models, index choices, similarity thresholds, and per-document matching with self-matches excluded.

Multi-Signal Integrity Detection

This branch covers the three integrity signals: internal peer similarity, external plagiarism against a local reference base, and AI-content likelihood, combined into one structured breakdown. Its children cover each signal and the discount applied to shared boilerplate. Keeping the signals separate prevents one noisy measure from driving the verdict.

Citation and Explainability Engine

This branch covers audit reports in which every score or penalty points to exact passages and verifiable locations. Its children cover citation pairs, citation metadata, JSON and Markdown report formats, justification rules, and fast cross-verification. A penalty without a citation is not an acceptable result.

10,000 Document Scaling Brief

This branch covers the architecture write-up for moving from the sample batch to 10,000 submissions. Its children cover compute choices, asynchronous queues, index strategy, and throughput bottlenecks. The brief is judged on the reasoning behind trade-offs rather than on a production-grade system.

Sample Submission Dataset

This branch covers building the sample_submissions corpus with the required file types, deliberately planted cases, and a ground-truth manifest. Its children cover corpus composition, planted cases, and the manifest format. Without known answers, the evaluation cannot measure anything.

Test Scenarios

This branch covers the five end-to-end scenarios the pipeline must pass on the sample data: cross-modal duplication, paraphrase detection, corrupted input handling, benign boilerplate, and multi-document batch evaluation. Its children describe the inputs and expected behavior of each scenario. Passing them is the clearest evidence that the pipeline works beyond the happy path.