I built Fanoni Clinical AI to make the path behind an answer visible: choose a focused skill, retrieve from the right source, and stop when the record or evidence cannot support the claim.
Choose a use case to see the tools and rules behind the response. These are synthetic prompts based on the repository configuration; the screen does not display patient records or a live agent answer. The guideline example uses coded rules in the repository, not a live guideline feed.
“What medications is Nikolai Chen taking?”
search_fhir_patientsget_patient_fhir_resourceFHIR chart dataNo match means stop. Multiple matches mean ask the clinician to choose.
Interactive portfolio walkthrough based on the skills configuration. The source repository is temporarily private during credential rotation.
The current skills configuration connects a NeMo router to four tool-calling agents. Each branch has a small, explicit tool list and an instruction set for what to do when data is missing, ambiguous, or unavailable.
The EHR agent surface sends a question through the clinical API to the NeMo service. The configured router selects one specialist skill instead of exposing every tool at once.
router_agent → one of four skillsmedical_ehr/configs/medical_ehr_SKILLS_gemini.ymlThe patient skill gets FHIR reads and cohort tools; medication safety, coding, and decision support each get their own shorter list. Each skill also has its own answer and failure rules.
skill → curated MCP / native toolsmedical_ehr/configs/medical_ehr_SKILLS_gemini.ymlThe MCP server reads the selected FHIR backend for chart data. Medication checks inspect drug labels, code searches use terminology tools, and named risk scores run through calculator functions.
MCP → FHIR + specialist sourcesclinical-mcp-server/tools/fhir_tools.pyA chart question first calls patient search. No match stops the flow; several matches are returned for a clinician to resolve before a record is read.
clinical-mcp-server/tools/fhir_tools.pyFHIR cohort tools count unique patients and support condition intersections and exclusions. The skill asks for numerator and denominator when it reports a percentage.
clinical-mcp-server/tools/fhir_tools.pyOptional Graphiti memory is keyed by patient ID. The skill instructs the agent to verify current clinical state against FHIR, and a memory outage should not block a chart answer.
medical_ehr/src/nat_medical_ehr/tools/graphiti_memory_tool.pyThe guideline tool matches a condition and proposed plan against coded rules. Literature lookup is a separate PubMed tool. The portfolio does not present either as a continuously updated clinical guideline service.
clinical-mcp-server/tools/tier2_tools.pyClinical questions span patient charts, medication references, coding sources, and published evidence. A general response is hard to trust when it skips patient matching, fills in absent records, or obscures which source was consulted.
This project gives each question a narrower route and a clear stopping rule. It is designed to help a clinician inspect information, not to make a diagnosis or treatment decision on their behalf.
Find a patient, read chart resources, or answer a cohort question from the FHIR store.
Check an interaction, label, boxed warning, or medication list through curated safety tools.
Look up ICD-10, CPT, or HCPCS data and review code or payment details returned by tools.
Compare plans with coded guideline rules, retrieve PubMed literature, or run deterministic scores for clinician review.
The clinician suite asks for an ambiguous surname and for patients absent from the test store. It checks that the agent asks for a choice or stops instead of inventing a medication list.
The same script checks selected medication, encounter, diagnosis, and population answers against a mirrored synthetic FHIR dataset. It also checks that counts describe unique patients.
Guideline and interaction prompts are checked for relevant source language, while an unsupported reimbursement question should say the data is unavailable. These checks are internal system tests, not clinical outcome evidence.
medical_ehr/evals/doctor_suite.pyThese Fanoni Hugging Face assets document earlier medical EHR training and structured-output research. The current pilot configuration uses Gemini 2.5 Flash through Vertex AI.
Synthetic EHR questions with expected answers, routing strategies, and terminology fields for earlier DSPy and GEPA experiments.
Open dataset ↗PUBLIC RESEARCH ARTIFACT / APACHE 2.0A published structured-output optimization artifact from earlier research. The current skills pilot uses Gemini on Vertex AI instead.
Open model card ↗The repository’s clinician-flow suite checks patient disambiguation, missing-record stops, chart answers, cohorts, medication interactions, and evidence routing. These are internal system checks. The project remains a pre-deployment pilot, and the portfolio walkthrough uses synthetic questions rather than a live patient record.