Grounded RAG & LLM systems
RAG pipelines and agents that cite their sources and refuse when the evidence runs out. I build new systems end to end and integrate AI into the workflows and tools you already have.
Biomedical AI · Knowledge Graphs · LLM Systems
I build biomedical AI
that shows its work.
I build LLM and deep learning systems that teams can trust: grounded retrieval, honest evaluation, and knowledge graphs that keep models accountable. Biomedical data is my specialty, and the methods carry anywhere. Before AI I spent a decade in frontline medicine. Today I’m an AI PhD researcher and a key maintainer of ARAX, a reasoning system in NIH’s Biomedical Data Translator.
562k-chunk clinical corpus · 7.5M-node knowledge graph · AIME 2025 · NCATS Translator
NCATS Translator
Key maintainer of ARAX, NIH’s biomedical reasoning agent
AIME 2025
Published at Springer’s AI in Medicine conference
Ramsey Lab, OSU
PhD researcher, AI · concurrent M.S. in CS
10 years in medicine
Army combat medic · paramedic · Cleveland Clinic · FEMA
Services
What I do
LLM evaluation & safety
Eval harnesses, faithfulness checks, adversarial testing, and data safeguards for systems you already run. You get numbers you can defend in a review.
Knowledge graph engineering
Schema design, ontology mapping, identifier normalization, and graph ML at the scale of millions of nodes. This is the plumbing that makes grounding possible.
Deep learning for prediction
Custom models for classification, ranking, and link prediction, built with interpretability in mind. My research focus is models that can show why they predicted what they did.
I also take on technical writing, from whitepapers to developer documentation. If your project sits near this list but not on it, ask anyway.
Engagement models and pricing →Selected work
Case studies
Sole builder · deployed on AWS · published at AIME 2025
Radiant: clinical evidence RAG for infectious disease
A decision support assistant for licensed clinicians. It answers dosing, isolation, and treatment questions only from curated evidence. Every claim carries a verbatim citation, and when the corpus doesn’t cover a question it says so instead of guessing.
I built it end to end. The corpus holds around 562k chunks from PMC full text, CDC, WHO, and IDSA guidelines, and institutional SOPs behind a license gate. Retrieval is hybrid dense plus BM25 with reranking. A three-tier evidence hierarchy weighs institutional guidance against national guidelines and research, with an explicit verdict when sources disagree. A PHI guard blocks identifiers before anything reaches an LLM. The system runs on AWS behind auth with encrypted secrets, and an adversarial test suite covers prompt injection and dangerous dose bait.
- 0.82 → 1.00 faithfulness on versioned gold-set eval, after a regenerate-then-refuse gate
- ≈0.9 behavior accuracy
- 32-way concurrent retrieval, zero errors
- Passed prompt-injection & dosing-bait adversarial suite
Key maintainer · NIH NCATS Translator · in production
ARAX: machine learning inside an NIH reasoning system
I’m a key maintainer of ARAX, the reasoning agent that plans multi-hop queries over the 7.5M-node RTX-KG2 knowledge graph. One long-standing weak point was overly generic nodes like “protein” and “disease” polluting reasoning paths, held back only by a hand-maintained blocklist. I replaced that blocklist with a learned classifier.
The pipeline extracts graph features and node embeddings over KG2, feeds an information content regressor, and trains an XGBoost classifier on top. An LLM-in-the-loop active learning stage running on a self-hosted vLLM server grew the labeled set from 11.5k to 127k nodes over four rounds. Everything was validated against the ARAX regression test suite before adoption.
- 0.961 PR-AUC
- 0.90 / 0.88 precision / recall at the deployed threshold
- 21,689 candidates triaged by LLM review in 757s (~29/s)
- 11.5k → 127k labeled nodes via active learning
Single-author research · methodology
CCLB: link prediction that is faithful by construction
When a model proposes a molecular mediator for wet-lab validation, a post-hoc explanation isn’t good enough. CCLB trains a hard-gated edge selector jointly with a Neural Bellman-Ford predictor, so every prediction is computed on a sparse subgraph the model is structurally committed to.
Trained on the human subset of the Reasoner Drug Knowledge Graph, 856k entities and around 2.7M triples, on DGX H100 hardware. The architecture meets all four of Azzolin et al.’s (ICLR 2025) conditions for self-interpretable graph models, audited condition by condition in the paper. My current work extends this line toward query-conditional faithfulness guarantees.
- 79% of unconstrained baseline MRR…
- ~12% …using this fraction of input edges per query
- 16× mean-rank reduction over training (26,395 → 1,614)
- 4 / 4 architectural faithfulness conditions met
Paper available on request.
More work
- PloverDB Maintainer In-memory graph query service powering Translator knowledge providers. I own deployment pipelines, build validation, and performance.
- localBabel Author Biomedical identifier and category normalization into a local SQLite schema. Tested, benchmarked, and published to PyPI.
- RTX Uptime Monitor Author Liveness and latency probes for production reasoning endpoints, with a live dashboard and Slack alerting.
- RareTarget Contributor Knowledge-graph framework for prioritizing intermediate molecular targets in rare disease.
- Node Embedder Author Embedding pipeline over KG node descriptions enabling semantic search on biomedical relationships.
- NL → Query Graph Author NLP/NLI pipeline translating natural-language biomedical questions into formal query graphs.
Publications & talks
On the record
-
F.M. Hodges et al., “Using AI to Improve Diagnosis and Treatment of Rare Diseases: A Chat Agent for Equitable and Accessible Healthcare,” Artificial Intelligence in Medicine (AIME), LNCS vol. 15735, Springer, 2025. doi:10.1007/978-3-031-95841-0_35
-
F.M. Hodges et al., “White Paper on Radiant,” Oregon State University, 2025. radiant.rtx.ai/whitepaper
-
Poster — Radiant: An Agentic RAG System for Rare Diseases. Artificial Intelligence in Medicine, Pavia, Italy, 2025.
-
Demo — Radiant. Stanford & Research to the People Rare Disease Hackathon, San Francisco, 2024.
About
Both sides of the problem
I spent a decade in medicine before I ever trained a model: combat medic in the Ohio Army National Guard, paramedic, medical assistant at the Cleveland Clinic, medic contractor with FEMA. I went back to school for computer science because I wanted to affect healthcare at a systems level. Now I work on machine learning and biomedical knowledge graphs as a PhD researcher in the Ramsey Lab at Oregon State University, with a concurrent M.S. in computer science.
That history shapes how I build. I’ve been the person relying on the protocol at 3 a.m., so I care less about leaderboard scores than whether a system can show its evidence, admit uncertainty, and fail safely.
Off the clock: cooking, my dog, and a Cleveland sports loyalty that no amount of evidence has managed to correct.
Contact
Let’s talk
I take a small number of consulting engagements alongside my PhD: scoped builds, evaluations, advisory, and technical writing. Tell me what you’re working on and I’ll reply within two business days.