Jonathan Wang

MetaFabric AI: Mission-Driven Data Discovery for the AI Era Data drives intelligent decisions, but data search hasn't kept up. Keyword search leaves users with seven problems: 1. It depends on perfect metadata, yet standards vary and fields are often incomplete. 2. It matches strings, not the meaning of titles, descriptions, or data dictionaries. 3. It can't turn a real-world mission into the data needed to address it. 4. Users must guess which keywords agencies use. 5. Natural-language questions don't work. 6. Nothing explains how datasets relate or how to combine them for analysis. 7. Data gaps stay invisible until deep into an analysis. Example: DC's open-data portal returned no results for "I am looking for data to support traffic safety analysis near DC schools." Shortened to keywords, it returned only two, both describing the same speed-camera data. How MetaFabric AI solves these problems Meta's Muse AI models understand both what the user wants to accomplish and what each dataset actually means, then match the two. - Mission-driven discovery (problems 3-5): Users describe a goal in plain language. For the same DC question, MetaFabricAI found 20 relevant datasets spanning schools, crashes, the High Injury Network, Safe Passage Routes, Vision Zero, violations, and traffic volume. - Semantic metadata understanding (problems 1-2): Muse reads names, descriptions, summaries, and categories for meaning, so incomplete metadata can still be found. It also makes keyword search smarter: "healthcare" finds hospitals, nursing homes, pharmacies, and clinics. - Explainable recommendations: Each result shows a relevance score, its role in the mission, why it fits, caveats, and a live source link. An AI Assistant refines results through conversation. - Data Recipes (problem 6): A structured plan with the analytical objective, dataset relationships and joins, a step-by-step workflow, and expected outputs. - Data gap analysis (problem 7): Each recipe lists gaps, assumptions, and limitations before analysis begins. How it's built - MetaFabricAI combines a React app, a FastAPI backend, and three Muse Spark agents (Discovery, Assistant, Recipe). Muse Spark's long context window lets it evaluate every catalog record, not a keyword-filtered shortlist. - MetadataMCPServer gives AI agents access to 1,660 metadata records through four cloud-hosted MCP tools. - OpenDataHarvester collects and standardizes metadata from government portals into one unified catalog. MetaFabric is metadata-first: data stays with its providers, access rules are respected, and every recommendation traces back to a real source. Future vision Next steps include authorized metadata discovery, publishing unmet data needs to connect users with providers, learning from past Data Recipes, and automatically combining discovered data to produce AI generated analysis reports.