RAPHAEL
MALIKIAN
Healthcare founder/operator โ I bootstrapped and ran a direct primary care clinic as CEO, and now I build open-source clinical AI infrastructure and publish open-access meta-analyses (ambient AI scribes, LLM citation fabrication) to fix the systemic pain points in healthcare. Over 20 years of healthcare domain expertise, shipped as code: products with AI built into the platform itself, not bolted on.
*Merged into external projects, verified live on GitHub Nov 15 2026 โ every one reviewed by that project's maintainers. Full record in the Field Notes.
RESEARCH
13 WORKS ยท UPDATED NOV 2026Open-access clinical AI research from an independent researcher: five Cochrane-style systematic reviews & meta-analyses, two evidence commentaries, and a six-paper open-source research series. Every work is a free Zenodo preprint (CC BY 4.0) with open data. Topics: ambient AI scribes, LLM hallucination & fabricated citations, clinical NLP de-identification, FHIR interoperability, vector databases & RAG, medical imaging reliability, direct primary care outcomes. No industry affiliation, no conflicts of interest.
Ambient AI Scribes Meta-Analysis โ Documentation Time, Burnout & Note Errors
Systematic review & meta-analysis of ambient AI scribes (AI medical scribes) for clinical documentation: 1,049 records screened, 74 studies included. Pooled documentation time fell 18.9% (low certainty) โ and the AI note-error burden is pooled alongside the benefit, so clinician review stays part of the intervention. Submitted to the Journal of Clinical Epidemiology.
LLM Citation Fabrication โ How Often AI Models Invent References
Systematic review of large language model citation accuracy in biomedical writing: 63 studies, 148 model arms. Pooled 19.4% of LLM citations were fabricated and only 37.0% fully accurate; retrieval grounding (RAG) cut fabrication to 5.8% vs 31.0% without. The reference-checking evidence base for clinical AI safety.
LLM Clinical Summarization Errors โ Omissions, Inaccuracies & Fabrications per Summary
Systematic review & meta-analysis of errors in LLM-generated clinical summaries across all summary types, with discharge summaries as a subgroup. Headline finding: the primary outcome could not be pooled because only one study in the literature reports it, and the 31 studies that could not be rescued for missing per-summary counts are presented as a formal reporting-gap result, not dropped quietly.
Unsafe Advice from Patient-Facing LLM Chatbots in Primary Care
Systematic review & meta-analysis of unsafe advice from patient-facing LLM chatbots in primary-care presentations: 3,106 records screened, 166 studies included. Pooled 10.5% of bench-evaluated responses fell in an unsafe tier (552 of 4,868; very-low certainty). Human-user studies were analysed separately and never pooled with bench work. Journal submission to JAMIA to come.
High-Dose Vitamin C in Cancer: What the Trials Actually Show
Systematic review & meta-analysis of high-dose vitamin C in cancer: 9,197 records screened, 18 studies included (10 RCTs). The intravenous overall-survival pool (HR 0.73) was judged uninterpretable at Iยฒ 87.5%, so the paper reports trial-level results instead: Wang/VITALITY 1.04, Bodeker 0.46, Ou 0.33, Paller 1.98. Journal submission to the Journal of Clinical Epidemiology to come.
Direct Primary Care Outcomes: 83% Growth, 38 Papers
DPC outcomes commentary: direct primary care grew 83% on an evidence base of just 38 outcomes papers โ the outcomes accounting the DPC model owes patients and payers. Healthcare MS HJDSI-D-26-00372; DPC Summit 2027 session proposed.
When a Warning Advertises โ 7-OH & the Boomerang Effect
Public-health communication commentary: anti-drug warnings can advertise the drug (boomerang / reactance effect), applied to concentrated 7-hydroxymitragynine (7-OH) products. Evidence & policy only โ no clinical recommendations. 62 references.
Guarding the Clinical Narrative
Clinical NLP & health-data de-identification: retrospective analysis of open-source contributions to clinical natural language processing and HIPAA de-identification (2025โ2026).
Silent Failures in the Imaging Stack
Medical imaging AI reliability: DICOM geometry, warnings discipline and silent pipeline failures in open-source medical imaging software (2025โ2026). MS ARAD-S-26-05599.
Retrieval Is the New Record System
Vector databases & embedding infrastructure for clinical AI and RAG: retrospective analysis of open-source contributions (2025โ2026), 39 PRs in this domain alone. MS JBI-26-4036.
The Plumbing Layer
FHIR interoperability infrastructure: open-source FHIR servers, synthetic patient data and federated research infrastructure โ the layer every clinical AI claim depends on (2025โ2026).
When Agent Plumbing Fails Quietly
LLM tooling & AI agent infrastructure: retrospective analysis of open-source contributions to agent frameworks โ the quiet failures that break AI pipelines downstream. MS CMPB-D-26-08971; FOSDEM 2027 talk proposed.
Synthetic Patients and What They Can and Cannot Teach
Synthetic patient data from Synthea: what synthetic patients can teach clinical AI, and where they mislead โ grounded in 7 open-source Synthea PRs, 6 merged. Journal submission to Health Information Science and Systems to come.
The earlier record
Peer-reviewed since 2009: four-station knee simulator study (J Bioengineer & Biomedical Sci 2013), knee arthroplasty roughening (The Knee 2014), JBJS proceedings, ORS & PRAS presentations.
THE DIGITAL BUSINESS CARD
MOVE YOUR CURSOR OVER IT ยท TAP TO FLIP THE DETAILSThis page is my business card. Scan it, save it, send it. The QR on the right drops my full contact straight into your phone โ no app, no signup.
RAPHAEL MALIKIAN
Building open-source AI for patients, providers & the systems between them.
Scan to save my contact โ name, email, GitHub, X & LinkedIn in one tap.
GET STARTED
Three doors, pick yours. Every one lands in my inbox or on my GitHub within a click โ I read everything myself.
I'm hiring / building clinical AI
You need someone who has run a clinic and ships production AI โ evals, safety harnesses, FHIR, OpenAI-compatible platforms that self-host for PHI. That's the whole portfolio below.
I want to collaborate on open source
I contribute daily to medical AI & vector DB projects and ship a big new open-source product every week. Open an issue, propose a feature, or point me at a bug you need fixed.
I have a problem AI should fix
Patient pain point, provider workflow, denial letters, discharge chaos โ if it's high-demand and underserved, it's probably already on my build list. Tell me yours.
AI VENTURES HUB
My signature as an architect: every flagship product embeds its own OpenAI-compatible AI endpoints (/v1/chat/completions ยท /v1/embeddings) with pluggable backends โ local Ollama/llama.cpp for PHI-safe self-hosting, or any compatible upstream โ plus a deterministic offline fallback. The AI is the platform, not a plugin.
โข CLINICAL AI โ WEEKLY PRODUCT LAUNCHES
Discharge Decoded
Photograph discharge papers โ a plain-language care plan: med schedule, red flags, follow-up checklist. Verified extraction (unverified meds are flagged, never presented as fact), a 6th-grade readability gate, and automated teach-back. Week 1 flagship, shipped Oct 19 2026.
Clinical LLM Safety Harness
LLM evaluation & safety harness for clinical AI: hallucination detection, citation verification, human-in-the-loop escalation โ the guardrails layer safe healthcare LLMs are missing.
AI Medical Clinic + Workflow Skills
Agentic medical-office automation: EHR/FHIR R4 utilities, SMART on FHIR workflow skills, embedded inference gateway for scheduling, intake & admin in small clinics.
ClaimGuard AI
Insurance denial management AI: predicts denial risk, finds the cause in the denial letter, drafts the appeal โ LLM distilled from 10,000 synthetic denial/appeal pairs, human-in-the-loop.
Azat โ Quit Smoking, Without Shame
ิฑีฆีกีฟ means โfree.โ An evidence-based, shame-free quit-smoking PWA in English, Russian and casual Eastern Armenian. No login, no data leaves the phone โ built for the Armenian diaspora, designed against fear/shame โboomerangโ effects.
Weekly Research Series
Nine open-access papers (Oct 2026): two Cochrane-style meta-analyses (ambient AI scribes; LLM citation fabrication), two commentaries (direct primary care outcomes; 7-OH boomerang effect) and a five-paper open-source series โ clinical NLP, medical imaging, vector databases, FHIR, LLM agent infrastructure. Zenodo DOIs, CC BY 4.0.
โข TESLA AI SERIES โ 22 APPS SHIPPED ยท ONE A DAY (OCT 2026)
22 open-source Tesla apps shipped Oct 8โ29, 2026. Series 2 (12 apps) embed OpenAI-compatible AI endpoints (/v1/chat/completions ยท /v1/embeddings), a pluggable LLM backend and a deterministic offline fallback โ each with a live demo. Series 1 (10 apps) are deterministic Tesla Fleet API automations.
Tesla Girlfriend Mode
Passenger comfort profiles for Tesla: weather-chosen climate & seat heating for the passenger seat, freeway window override, Spotify mix on activation โ chat-activated AI endpoint with on-device learning. Bonus build, Oct 29.
Tesla Climate Concierge
LLM calendar parsing for Tesla preconditioning: free-text events (โearly flight, leave 5ishโ) become structured departure plans with weather-aware cabin preconditioning and Cabin Overheat Protection re-arm.
Tesla Trip Brain
Tesla trip-readiness planner: calendar + battery state of charge + destination weather โ required departure charge and preconditioning plan. The LLM explains; the rules engine computes.
Tesla Drain Forecaster
Phantom drain prediction for parked Teslas: learns each car's overnight vampire-drain pattern, predicts tonight's loss and the likely culprit, with chat Q&A and embedding search over past nights.
Tesla Battery Guru
Tesla battery health analytics: degradation and usable-capacity estimates from charging-session data, charging-habit scoring, and an AI coach endpoint grounded in your computed battery stats.
Tesla Charge Ledger
Free Tesla charging-cost ledger: every session priced against real time-of-use (TOU) electricity rates, monthly reports, CSV export, and natural-language cost Q&A via its AI endpoint.
Tesla Family Guard
Watch-only Tesla teen-driver safety monitor: speeding, geofence and curfew alerts, plus a weekly AI driving report card per driver. Never commands the car.
Tesla Rain Guard
Automatic Tesla window closing before rain: weather radar at the parked car's location closes vented windows via the Fleet API; the AI endpoint explains every decision from its log.
Tesla Mileage Ledger
IRS-ready Tesla mileage log from exact odometer data โ no $60โ120/yr GPS app: automatic trip classification, business-mileage deduction math, CSV/HTML exports, AI trip-purpose suggestions.
Tesla Home Link
Tesla-to-smart-home automation: ETA-aware thermostat pre-heating/cooling (Nest, Ecobee, Home Assistant) and Hue lights on approach, with a chat endpoint that explains the trigger log.
Tesla Location Charge Limits
Geofenced per-location Tesla charge limits (home 80%, cabin 90%) โ the Rivian feature Tesla lacks โ set once per arrival, with natural-language charging rules via AI endpoint.
Tesla Fleet Chat
Chat with your Teslas from any OpenAI client: ask โwhich car has more battery?โ or run allowlist-validated Fleet API commands. OpenAI-compatible endpoints, offline rules fallback.
โข TESLA SERIES 1 โ FLEET API AUTOMATIONS
Tesla Charge Coordinator
Multi-Tesla home charging coordination: priority queue so only one car charges at a time, off-peak windows, automatic handoff. Pre-series standalone, Oct 8.
Tesla Calendar Climate
Calendar-aware Tesla preconditioning: reads your iCal, preconditions the cabin with weather-aware lead times, re-arms Cabin Overheat Protection before its 12-hour cutoff.
Tesla Sentry Scheduler
Sentry Mode on a real schedule (default 22:00โ06:00): cuts overnight phantom drain without losing nighttime security.
Tesla Drain Detective
Overnight vampire-drain alerts that name the culprit โ Sentry, climate keeper or Cabin Overheat Protection โ from no-wake reads only.
Tesla COP Sentinel
Keeps Cabin Overheat Protection alive past Tesla's 12-hour cutoff in parked heat by re-arming it automatically (COP = cabin overheat, not police).
Tesla Valet Auto-Arm
Geofenced Valet Mode auto-arm at car washes, service centers and valet stands, with instant owner alerts.
Tesla Dog Mode Guardian
Backup Dog Mode safety alerts to a second contact if cabin temperature drifts or battery falls low. Alert-only, never a life-support replacement.
Tesla GlareGuard
One-tap anti-glare mirror deflection: computes exact mirror angles from seat position and driver height to steer headlight glare below the eyes; auto-restores in 60s.
Tesla Peak Pause
Time-of-use peak charging pause: stops charging at peak start (e.g. SCE 4โ9pm) and resumes off-peak โ the end time Tesla's own scheduler lacks.
Tether (macOS)
Evidence-informed screen-wellness menu-bar app for macOS: 20-20-20 break overlays, movement breaks, Pomodoro, privacy-first AI-usage awareness โ โsees when, never what you type.โ
REPOSITORY INDEX
LIVE FROM THE GITHUB APIEvery original product, pulled live from GitHub โ this list never goes stale. My 95 contribution forks (Ollama, Chroma, Milvus, txtai, OpenMed & more) live on my GitHub profile.
Loading live from GitHubโฆ
RรSUMร SNAPSHOT
The short version. The long version is the work itself โ it's all public, linked, and verifiable above.
Clinical AI Architect ยท Open-Source Builder
Weekly clinical AI product launches with embedded OpenAI-compatible inference gateways; daily contributions to medical AI & vector DB projects (243 PRs opened, 73 merged into external projects); nine open-access research papers in Oct 2026 (two meta-analyses, two commentaries, five-paper series); 22 Tesla AI apps shipped in a daily series. Volunteer literacy/math tutor, Palmdale City Library; taught a 3-hour community Computer & AI masterclass.
Founder & CEO โ Direct Primary Care Clinic
Bootstrapped and ran a direct primary care clinic as CEO: operations, team, finances, patient experience. This is where the pain points in my products come from โ I lived them.
Family Medicine Residency โ Eastern Virginia Medical School
Residency training in family medicine, with a podium presentation on renal-insufficiency prescribing (EVMS 2018).
ECFMG Certified
ECFMG certification verifying my UCL medical degree for US training and practice pathways.
UCL โ MBBS (2011) ยท BSc First Class Honours (2008)
University College London. Dean's List 2007/08. Published researcher since 2009: knee arthroplasty & simulator work in The Knee (2014), J. Bioengineering & Biomedical Science (2013), JBJS proceedings and ORS.
STACK & SIGNALS
SELECTED PUBLICATIONS
2026: Ambient AI scribes meta-analysis (submitted, J Clinical Epidemiology) ยท LLM citation fabrication meta-analysis ยท Direct primary care outcomes commentary (submitted, Healthcare) ยท Five-paper open-source series: Clinical NLP & De-identification ยท Medical Imaging ยท Vector Databases ยท FHIR Infrastructure ยท LLM Agent Infrastructure (all Zenodo, CC BY 4.0) ยท Four-station knee simulator study, J Bioengineer & Biomedical Sci 2013 ยท Retrieved knee arthroplasty roughening, The Knee 2014 ยท JBJS Br proceedings 2010/2012 ยท ORS 2011 ยท PRAS 2009.
SIGNAL LOG โ BLOG & X
Two streams, one page: Field Notes โ balanced reflections on what we're building and contributing, including what didn't land and why โ and From X, my recent posts and threads. Follow @rtmalikian for the live version.
LET'S BUILD SOMETHING
I'm open to clinical AI architect roles, collaborations, and good problems. The fastest way to reach me is email โ I answer personally. Or grab the card above and keep me in your pocket.