Yana Zalewska

2025-01

Failure mode knowledge graph for AI FMEA assistance — built a structured graph of 2,400 aerospace failure modes with relationships to affected components, detection methods, and mitigations, drawn from historical FMEA databases.

Challenge: Deduplication of failure modes across 12 FMEA source databases was the core challenge — nominally identical failure modes were described with different terminology across programs. Implemented a combined fuzzy text matching + embedding similarity deduplication that reduced 4,800 raw failure mode records to 2,400 canonical entries.

Learned: FMEA failure mode deduplication requires hybrid text + semantic matching. Pure fuzzy text matching misses semantically equivalent failure modes with different wording. Embedding similarity catches paraphrases but is too permissive for unrelated failure modes with similar surface forms — the combination of both approaches outperforms either alone.

2024-06

Graph RAG implementation — combined Neo4j knowledge graph traversal with Qdrant vector search for engineering document queries. Graph pre-filters the retrieval candidate set using requirement relationships before vector similarity ranking.

Challenge: Graph traversal for requirement relationship chains was slow for deeply nested requirement hierarchies (5+ levels). Required caching the traversal results for frequently queried top-level requirements, reducing graph traversal time from 800ms to 45ms for common queries.

Learned: Graph RAG requirement traversal performance depends critically on graph depth. For hierarchical engineering requirement structures, pre-compute and cache the transitive closure of derivedFrom relationships — re-traversing 5+ level hierarchies on every query is unacceptable for interactive RAG applications.

2023-11

Neo4j knowledge graph population — built an ingestion pipeline that extracts entities and relationships from engineering documents (PDFs, DOORS exports, FMEA spreadsheets) and populates the graph. Used spaCy NER + rule-based relation extraction.

Challenge: Relation extraction from FMEA spreadsheets was straightforward, but relation extraction from unstructured engineering PDFs had 65% precision — too many false relationships. Implemented a confidence threshold and human-in-the-loop validation queue for low-confidence extractions.

Learned: Knowledge graph population from unstructured engineering documents requires a human validation layer. NER + relation extraction precision on engineering text is typically 60-70% — sufficient to propose candidate relationships, but not reliable enough to populate the production knowledge graph without review.

2023-04

Aerospace engineering ontology design — defined core ontology concepts (Component, Document, Requirement, FailureMode, AnalysisResult) and relationships (hasPart, derivedFrom, satisfies, identifiesFailureOf) for the engineering knowledge graph.

Challenge: Aerospace engineering vocabulary has significant terminological inconsistency — the same physical component is called a 'part', 'assembly', 'item', or 'article' depending on the document source. Defining canonical ontology terms required aligning terminology across 4 engineering disciplines over 3 workshops.

Learned: Engineering ontology design cannot be done without multi-discipline terminology alignment workshops. Abstract ontology modeling produces an academically correct but practically unused graph — the ontology must be grounded in the actual vocabulary used by engineers in each discipline.