Capabilities
Why governance comes first
A clinical model that learns in deployment changes the system it was validated in. The firm's research program treats governance as a viability constraint on adaptive computation rather than a compliance layer added after the fact. Every capability below is built to that premise: generation separated from evaluation, thresholds declared at named stages, and an audit trail that survives the model that produced it.
Clinically consequential AI failures arise from interactions among workflows, documentation practices, incentives, and accountability, not from isolated algorithmic defects. Peer reviewed in Am J Biomed Sci & Res (2026).
Shortcut learning and misclassification pass standard validation undetected. Audited in JAMIA Open (Oxford University Press, 2026), Harvard GCSRT capstone.
Externally Governed Learning Systems (EGLS) gives adaptive computation a formal viability constraint. Foundation of US Provisional Patent 63/975,551 and the AIDD-GOV open standard.
AI competency map
Each capability is exercised inside at least one portfolio platform. Maturity is stated as observed, not as intended. Evidence links resolve to public preprints, peer reviewed articles, patent filings, or public repositories.
| # | Capability domain | Product instantiation | Evidence on record | Maturity |
|---|---|---|---|---|
| 01 | Governed AI architecture and strategy | FxMED OS governance ledger · DrugSynthAI stage gates | EGLS, SSRN 6160268 · Patent 63/975,551 | Production, internal |
| 02 | Multi agent clinical AI systems | DrugSynthAI · MedClaw runtime | 67 agents, 5 tiers, 536 tests · Patent 64/018,624 | Production, internal |
| 03 | Predictive modeling and validation audit | FHIR native governance · bandit model selection | JAMIA Open ooaf177 · Research Square rs.3.rs-8714776 | Peer reviewed |
| 04 | Imaging and multi omics analytics | Genomic subtyping audit · mitochondrial variant modeling | JAMIA Open · MitoCoreX Zenodo 19393450 | Research |
| 05 | Clinical NLP and document intelligence | FxMED OS AI scribe · literature surveillance pipelines | FxMED OS service layers · methods published | Production, internal |
| 06 | Retrieval grounded generation | MedBoardPRO Analyst, Strategist, Judge · citation verified education engine | Patent 64/012,574 · Trademark 99721498 | Beta |
| 07 | MLOps, evaluation gates, deployment governance | DrugSynthAI build discipline · AIDD-GOV | 103 numbered rules · 18 stage gates · Apache 2.0 standard | Operational, open |
| 08 | Companion and conversational AI | FxMED Advisor inside FxMED OS | Second EGLS instantiation, in development | Phase 1 |
Capability domains
Each domain follows the same contract: the question the firm is investigating, the product context in which the capability is exercised, and what has been placed on public record.
Use case. Architectural core of FxMED OS and DrugSynthAI. Governance declarations name every agent's authority boundary; a ledger (SOGDR) records each release decision; the Watcher Protocol audits runtime behaviour against the declaration.
Use case. DrugSynthAI: sixty seven autonomous agents across five tiers (pipeline, validators, orchestrators, intelligence, precision medicine), 167 API endpoints, 536 passing tests including 37 security focused. MedClaw provides the local first orchestration runtime with cost aware model routing, Ollama primary inference, and cloud fallback.
Use case. Adaptive model selection via multi armed bandits inside HL7 FHIR native governance infrastructure. Empirical audit of shortcut learning in AI based genomic subtyping, peer reviewed in JAMIA Open, Vol. 9, Issue 2 (DOI 10.1093/jamiaopen/ooaf177).
Use case. MitoCoreX, the first validated DrugSynthAI campaign: systems level mitochondrial pathway connectivity mapping, druggability assessment of structure function constraints, and variant modeling for functional defect classification. Published on Zenodo (DOI 10.5281/zenodo.19393450), ChemRxiv, and Research Square.
Use case. The FxMED OS AI scribe operates inside the encounter, clinical document, and consent layers under business associate agreement governance. Research infrastructure pipelines ingest, screen, and synthesise biomedical literature across PubMed, Scopus, Embase, and preprint servers on a continuous schedule.
Use case. MedBoardPRO's three tier Analyst, Strategist, Judge architecture under EGLS governance (US Provisional Patent 64/012,574, 27 claim specification; USPTO Trademark Serial 99721498). A citation verified education engine resolves every reference through Crossref before a module is accepted and rejects degraded identifier metadata.
Use case. DrugSynthAI build discipline under 103 numbered rules (R01 to R103). Three campaigns, 163 candidates ranked per campaign, 18 stage gate decisions, all passed, zero kill switch activations. The AIDD-GOV governance standard (v0.1, Apache 2.0, 10 formal schemas, 3 conformance levels) publishes the pattern for reuse.
Use case. FxMED Advisor, the second production instantiation of externally governed learning systems, embedded inside FxMED OS as an operator specific AI layer. Phase 1, in development.
Platform engineering
Governed AI is only useful inside a system that clinicians, learners, and researchers can operate. The firm builds that system layer itself: electronic health record infrastructure, learning platforms, cloud native delivery, and the data engineering underneath. Each domain is stated with the platform that exercises it and its observed maturity.
| # | Platform domain | Product instantiation | Evidence on record | Maturity |
|---|---|---|---|---|
| 09 | Digital transformation for clinical organizations | FxMED OS · guided setup, onboarding, white label, tenancy | 19 service layers in production | Production, internal |
| 10 | Electronic health record and physician operating systems | FxMED OS registry, scheduling, encounters, documents, consent | HL7 FHIR native · BAA governance · Patent 63/975,551 | Production, internal |
| 11 | Educational and learning platforms | MedBoardPRO · AFMI Platform · bilingual course engine | Patent 64/012,574 · Trademark 99721498 · 104 gated modules | Beta / MVP |
| 12 | Healthtech product engineering | Patient portal, communications, membership, CRM layers | FxMED OS service layers · FxMED Advisor | Production, internal |
| 13 | Cloud native delivery and DevOps | Vercel, Supabase, Docker, git integrated deploys; MedClaw local first runtime | Deploy content gates · 536 tests · reproducible environments | Operational |
| 14 | Data engineering and research analytics | Governance telemetry database · literature pipelines · open datasets | Zenodo and Harvard Dataverse releases · CC BY 4.0 | Operational |
Use case. FxMED OS ships guided setup, tenant scoped onboarding, white label configuration, entitlements, and task workflow as first class service layers, so that a clinic's existing process is encoded rather than replaced. Multi tenant architecture with tenant scoped access control.
Use case. FxMED OS, the firm's Intelligent Physician Operating System, runs nineteen service layers in production: patient registry, scheduling, queue flow, encounter management, clinical documents, consent and permissions, attachments and media, portal policy, communications, AI scribe, tenancy, task workflow, entitlements, search, onboarding, white label, proxy access, membership, CRM, and guided setup. Architectural core covered by EGLS Patent 63/975,551.
Use case. MedBoardPRO, an adaptive licensure preparation engine for USMLE and related boards, runs a three tier Analyst, Strategist, Judge architecture under EGLS governance (US Provisional Patent 64/012,574; USPTO Trademark Serial 99721498). The AFMI Platform is a physician first educational SaaS on Next.js and FastAPI with Clerk authentication, Supabase, and multi tenant billing. A bilingual course engine produced 104 gated continuing education modules in functional medicine and endocrinology, every citation resolved through Crossref.
Use case. Patient portal policy, communications, notification and reminder workflows tuned to adherence objectives, proxy access, membership, and CRM run as FxMED OS service layers over the same governance ledger. FxMED Advisor, in development, extends the same core with an operator specific companion layer.
Use case. Git integrated production deploys on Vercel with Supabase and PostgreSQL as the data layer, Docker for reproducible environments, and deploy content gates that block any release carrying a gated artifact. The MedClaw runtime keeps inference local first through Ollama with cost aware cloud fallback, so the same pipeline runs on a workstation or in the cloud.
Use case. Per agent telemetry is logged to a governance database for regression detection and post hoc review. Literature pipelines ingest, screen, and synthesise biomedical evidence on a continuous schedule with declared inclusion rules and temporal weighting. Research data is released on Zenodo and Harvard Dataverse under CC BY 4.0.
Showcase cases
Each case is a research instantiation built and operated by the firm. Problem, approach, and what is on record; no outcomes are claimed beyond what has been published or filed.
Problem. Clinical software accumulates AI features faster than it accumulates accountability. Approach. Nineteen service layers on one tenancy model, HL7 FHIR native, with a governance ledger and runtime watcher derived from the EGLS framework. On record. Production, internal; EGLS Patent 63/975,551; BAA governed deployment posture.
Problem. Adaptive tutors optimise for engagement and drift from the evidence. Approach. Analyst, Strategist, Judge tiers with declared thresholds; a multi tenant physician first SaaS; a bilingual course engine that rejects any unverifiable citation. On record. Beta and MVP; Patent 64/012,574; Trademark 99721498; 104 gated modules.
Problem. Rare genetic disease is deprioritised by conventional pipelines. Approach. Sixty seven agents across five tiers, 103 build rules, stage gates with kill switch. On record. Three campaigns, 163 candidates ranked each, 18 gate decisions passed; Zenodo 10.5281/zenodo.19393450; Patent 64/018,624; AIDD-GOV open standard.
Development methodology
Nothing is built without an explicit success criterion and a named way it can fail. The same sequence governs a drug discovery campaign, a physician operating system release, and a licensure preparation model.
Research question, data provenance, clinical and regulatory constraints, and the enumerated ways the system can fail, written before architecture.
Agent responsibilities, authority boundaries, thresholds, and key performance indicators declared in a governance document that the build is tested against.
Every external identifier resolved at source. Degraded metadata rejected. Training and evaluation sets separated by contract.
Release through stage gates with kill switch protocols. Integration through standard APIs and HL7 FHIR where clinical data is in scope.
Runtime behaviour audited against the declaration. Regression detected from per agent telemetry. Retraining is itself a gated release.
Domains of study
The firm's research is concentrated where a practicing endocrinologist and clinician scientist can direct, audit, and clinically review the output. Domains are listed by the platform or program in which they are active.
| Domain | Program or platform | Representative output |
|---|---|---|
| Drug discovery for rare genetic disease | DrugSynthAI · MitoCoreX | Validated campaign, Zenodo 19393450; ChemRxiv preprint; AIMed 2026 Krakow poster |
| Clinical operations and physician workflow | FxMED OS | Nineteen service layers in production under BAA governance |
| Medical education and licensure | MedBoardPRO · AFMI Platform | Adaptive preparation engine, beta; educational SaaS, MVP |
| Genomics and precision medicine | Validation audit program | JAMIA Open 2026; Auditable AI for Genomic Equity, SSRN 6159546 |
| Cardio endocrinology and metabolic medicine | Clinical research program | 25 year systematic review, Int. J. Cardiovascular Medicine 2024 |
| Clinical AI governance and policy | EGLS · AIDD-GOV | Formal theory on SSRN; open standard under Apache 2.0; ACMG 2027 invited talk |
Technology in use
Chosen for reproducibility and local first operation. Model providers are interchangeable behind the orchestration gateway; governance does not depend on any single vendor.
Python · TypeScript · SQL
PyTorch · scikit-learn · RDKit · REINVENT4 · Boltz
Anthropic · OpenAI · Ollama (local first) · cost aware routing via MedClaw
Next.js · FastAPI · Supabase and PostgreSQL · Clerk · Vercel · Docker
HL7 FHIR · ES256 JWT · REST
AIDD-GOV schemas · governance ledger (SOGDR) · Watcher Protocol · Crossref verified citation engine
Operating principles
Each principle corresponds to a mechanism in the build, not to a statement of intent.
Every platform is directed, audited, and clinically reviewed by a practicing board certified physician. Clinical review is a named stage gate.
Adaptive computation is permitted only inside declared bounds. The bound is enforced by a ledger and a watcher, not by policy text.
The agent that produces an output never grades it. Critic actor separation is a contract, tested in the suite.
Citations resolve through Crossref before acceptance. Degraded metadata is rejected. Nothing is transcribed by hand.
Per agent telemetry, stage gate decisions, and kill switch state are logged for post hoc review and regression detection.
AIDD-GOV is published under Apache 2.0 with formal schemas and conformance levels so the governance pattern can be audited and reused.
Scope and status
Governed artificial intelligence platforms for drug discovery, licensure preparation, and clinical operations, each one a research instantiation of the firm's program on externally governed learning systems. Five products and three US provisional patents are on record.
Pre seed research and development. Two platforms operate in production as research instantiations, one is in beta, and two are in earlier phases. The emphasis is on validated proof of concept, peer reviewed audit, and published methods.
Yes. FxMED OS is an electronic health record and physician operating system with nineteen service layers; MedBoardPRO and the AFMI Platform are learning platforms; the delivery, cloud, and data engineering underneath are built in house. The AI capability domains and the platform domains share one governance framework.
Through the EGLS framework: a governance declaration names each agent's authority boundary, a ledger records release decisions, a watcher audits runtime behaviour against the declaration, and stage gates with kill switch protocols control release. The pattern is published as the AIDD-GOV open standard.
By empirical audit rather than assertion. The firm's peer reviewed work on shortcut learning shows how misclassification evades standard validation; the platforms carry that audit into deployment with per agent telemetry and regression detection. Protected health information is handled only under an executed business associate agreement.
Anthropic, OpenAI, and local inference through Ollama, routed by cost and sensitivity behind the MedClaw gateway. PyTorch, scikit-learn, RDKit, REINVENT4, and Boltz on the learning and chemistry side. Next.js, FastAPI, Supabase, and HL7 FHIR at the application and interoperability layer.
On the Publications page: 31 papers on SSRN, 27 DOIs across platforms, peer reviewed articles in JAMIA Open and Am J Biomed Sci & Res, and an ORCID record with seventy verified peer reviews.
Scientific, editorial, and press correspondence is reviewed within one business day. Identify the journal, institution, or outlet in the subject line. Physicians and principal investigators receive direct founder contact.
Correspondence
The firm welcomes scientific and academic correspondence on any of the domains above, editorial inquiries, and press requests. Identify the institution or outlet in the subject line.