I'm Matthew Morris, a machine learning engineer with 15 years supporting national security missions and 9 building NLP and LLM systems. Through Aion Innovations I design retrieval pipelines and LLM workflows that stay grounded in their sources — document ingestion through embedding, retrieval, inference, and analytic delivery.
Roughly 765,000 open-source documents through an extraction pipeline — salience gating, LLM entity resolution, event clustering, materiality scoring — served as a React and FastAPI application with retrieval-backed research chat. Every figure drills back to the source documents behind it, with self-reported and third-party claims flagged separately.
Socratic tutoring for ages 8–13 that gives hints rather than answers. Moderation runs on both student input and model output, with daily parent summaries and safety alerts. An eval suite covering jailbreak attempts gates every prompt and model change before it ships.
A WordPress site rebuilt as a static Astro build for an owner who doesn't touch code. Episodes pull from the podcast RSS feed at build time, so the site stays current on its own; GitHub Actions deploys over SFTP to staging and production, with the original site retained as a rollback path.
RAG architectures end to end — embedding pipelines, vector indexing in pgvector, semantic search, and context strategies that hold up across large collections.
Ingestion, extraction, normalization, chunking, and semantic indexing — the unglamorous half that decides whether the retrieval above works.
Analysis that traces back to its sources, with evaluation in the loop so changes to a prompt or a model are caught before they ship.
Open to consulting and project work. Reach me at hello@aionbuilt.com.