Competitions

Extract entities and relationships from scientific abstracts to build a queryable knowledge graph. Input: 5,000 PubMed abstracts. Output: Neo4j-compatible graph with entities (people, institutions, diseases, treatments) and typed relationships. Must achieve > 85% precision on held-out extraction test.

Build a computer vision model that identifies manufacturing defects in electronic component photographs from assembly line cameras. The model must classify defects into 8 categories (scratch, crack, dent, discoloration, missing_pin, solder_bridge, correct, unknown) with at least 95% accuracy. Must be deployable on edge devices with < 50MB model size and run inference under 100ms per image on CPU.

Build an AI system that generates high-quality product descriptions in 5 languages (English, Spanish, French, German, Japanese) from a product data schema. Descriptions must be culturally appropriate, SEO-optimized, and consistent in tone. Output must be structured JSON with per-language copy.

Build a streaming anomaly detection system for industrial IoT sensor data. The system must process a continuous stream of 50 sensor channels at 10Hz, detect anomalies within 500ms of occurrence, and minimize false positives while maintaining high recall. Will be deployed on edge hardware with limited compute.

Given a corpus of 10,000 legal contracts, build a Retrieval-Augmented Generation pipeline that accurately answers natural language questions about specific clauses, obligations, and deadlines. The system must handle cross-document references, resolve ambiguous pronouns to correct entities, and provide exact source citations (document ID + page number) for every answer.

Build an automated pipeline that classifies business documents (contracts, invoices, purchase orders, NDAs) into their correct category. The pipeline must handle PDFs, extract text reliably, classify with high accuracy, and output structured JSON with confidence scores.