Senior Dental & Facial Surgeon (DDS, 18+ Yrs) | Clinical Product & Workflow Advisor
Clinical AI Evaluation (SOAP & PDSQI-9) • EHR Usability (<= 3 Clicks) • Conceptual HL7 FHIR & SNOMED CT Certified • AI-Assisted Product Discovery

I am a Senior Dental, Oral & Maxillofacial, and Facial Aesthetic Surgeon (DDS / Dentist, 18+ years of uninterrupted clinical and surgical practice; 5-Year First Professional Degree / EU EQF Level 7) working at the intersection of Clinical Product Advisory, Clinical AI Evaluation (Human-in-the-Loop / SME), and EHR Usability.

After nearly two decades in operating rooms and outpatient chairs, I grew tired of clinical software designed by people who never had to chart a surgical complication with gloves on. That frustration led me to initiate an independent clinical product discovery and applied R&D lab (templ.us), focused on reducing clinician documentation burden and catching silent AI hallucinations before they reach a medical record.

Transparency on My Technical Role: I am not a career software developer, database administrator, or traditional HL7 FHIR / data engineer. I am a practicing surgeon with foundational training in Health Informatics (SNOMED CT Foundation Certified, Oct 2026; conceptual HL7 FHIR R4/R5 architecture). I build and audit clinical software by operating as the Clinical Architect, Specification Lead, and Final Safety Auditor, orchestrating specialized multi-agent, multi-LLM engineering workflows in OpenCode under Andrej Karpathy's surgical execution principles.

Commercial LLMs are great at writing boilerplate code, and terrible at clinical nuance. Left unsupervised, a single agent drifts, mixes database schemas, or silently drops a negation particle in a post-op note. To solve that, I don't write the underlying TypeScript, Swift, or Python manually; instead, I run a closed-loop clinical engineering pipeline where distinct language models act as specialized subagents under strict verification gates:

1. Context Hygiene & Codebase Indexing (@explore): Fast-scanning agents map existing specifications, clinical dossiers, and repository invariants before a single line of code is planned.
2. Architectural & Schema Gate (@architect): High-reasoning models review state machines, multi-tenant PostgreSQL Row-Level Security (RLS) boundaries, and HL7 FHIR / SNOMED CT domain contracts prior to implementation.
3. Minimal-Blast-Radius Execution & Verification (@general & @surgeon): Coding agents implement strictly scoped changes and run automated typechecks, builds, and deterministic unit suites. If a fix fails twice, execution stops and escalates back to architecture.
4. Multi-Tenancy, Privacy & Clinical Safety Gate (@security-review & Independent Clinical Oracle): Every change touching authentication, encrypted PHI (AES-256-GCM), or clinical synthesis is audited by an isolated security reviewer and tested against an Independent Blind Clinical Oracle (PDSQI-9, O(M) rolling-row mWER, and AHRQ clinical harm rubrics).

While evaluating first-generation ambient scribes in live clinical care, I noticed they routinely fail on four clinical fronts: they omit ~18% of affirmed clinical facts, invert sentence-level negations, cross-contaminate RIGHT/LEFT anatomical sites in multi-site procedures, and leak treatment budget discussions into the legal clinical evolution.

To address those failure modes in my R&D prototype (Templus Scribe v6.2 Beta), I directed the multi-agent pipeline to build six mechanisms that set it apart from standard single-prompt wrappers:

• EHR-Agnostic Standalone Utility (Zero Lock-In): Operates as a lightweight macOS Menu Bar native utility (335px HUD, Universal Binary), a Web Workspace, and a Mobile Pocket Recorder (Screen Wake Lock) with a 2-pill zero-training interface ([Clinical Evolution] and [Transcript]), decoupled from any specific EHR database and enforced by 72-hour ephemeral retention with 1-click Crypto-Shredding.
• Edge DSP + Pure 2-Stage Acoustic-to-Clinical Split: Client-side audio conditioning (16 kHz anti-aliased decimation, 80 Hz high-pass pre-emphasis for surgical/dental ambient noise, and moderate dynamic P10 Voice Activity Detection that preserves short negation particles like "no" or "sem") feeds a strict two-stage multi-LLM pipeline:
  • Stage 1 (Pure Acoustic Clean-Verbatim Diarization): Dedicated strictly to faithful multi-speaker transcription with temperature: 0.0 and thinkingBudget: 0, eliminating intermediate JSON tuple compression bottlenecks.
  • Stage 2 (Contextual Identity & 8-Discipline Clinical Synthesis): Maps speaker identities (CLINICIAN, ASSISTANT, PATIENT, COMPANION) and structures notes tailored to 8 regulated healthcare professions (Medicine, Dentistry/OMFS, Nursing, Physical Therapy, Occupational Therapy, Nutrition ADIME, Psychology, and Speech Therapy) with thinkingBudget: 0 to prevent reasoning models from extrapolating unperformed normal physical exams.
• Six Deterministic Sub-Millisecond Code Guardrails (< 1 ms): Before a note is shown to the clinician, deterministic code verifies (1) Role-Authority & Temporal boundaries, (2) Sentence-Level Negation & Uncertainty scope, (3) Numeric Dosage Conservation (mg, mcg, mL, UI, FDI tooth numbers), (4) Anatomical Laterality Cross-Talk (Right vs. Left), (5) Commercial/Financial Pricing Scrubbing from the medico-legal evolution, and (6) Safe Patient Identity Anchoring (Zero Wrong-Patient Charting when companions or referring doctors are mentioned).
• Independent Blind Clinical Oracle (PDSQI-9 = 44/45, Fact F1 = 100.0%): A separate auditing service scores generated notes blindly against the ground-truth audio and transcript across 5 peer-reviewed clinical safety axes (PDSQI-9 instrument, CREOLA omission taxonomy, mWER, and AHRQ harm severity), achieving 44/45 (97.2%) on PDSQI-9 and 100.0% F1 recall/precision on affirmed clinical facts in a real 55-minute outpatient consultation with zero hallucinations or polarity inversions.

EHR-Agnostic Ambient Clinical Scribe (v6.2 Beta) & Hexagonal Multi-Jurisdictional EHR/ERP Blueprint (Core Spec v0.3.3)

• Architectural blueprints, Mermaid clinical dataflow diagrams, scientific benchmark summaries (PDSQI-9), and Architecture Decision Records (ADR-001 through ADR-007) for Templus (templ.us).
• Documents both the Active Standalone Ambient Scribe Track (ADR-007) and the Foundational Enterprise EHR/ERP Blueprint (ADR-001 to ADR-005, frozen reference) built around a 49-table PostgreSQL RLS Universal Core, 7 Jurisdictional Ports, 5 UI Extension Slots, and a 65-system Legacy EHR Migration specification.

Zero-Dependency Python Verification Library & Clinical Scribe Guardrails (MIT License)

• Open-source Python verification suite specified and clinically audited by me, and implemented through multi-agent AI engineering in OpenCode.
• Includes 13 deterministic validators and guardrails: SNOMED CT (SCTID Verhoeff dihedral $D_5$ check & v2.3.1 Compositional Grammar builder), SNOMED CT MRCM Role-Authority & Temporal Firewall, Sentence-Level Negation Verifier, Numeric Dosage Conservation Guard, Anatomical Laterality Cross-Talk Verifier, Medico-Legal Commercial Pricing Scrubber, Safe Patient Identity Extractor (Zero Wrong-Patient), US NPI (Luhn 80840), US DEA, FDA UDI (GS1), Brazil CNS, and HIPAA Safe Harbor PHI scrubbing.

• LinkedIn: linkedin.com/in/drtrentini
• Focus & Availability: Available for Clinical AI Evaluator / Medical-Dental Subject Matter Expert (SME), Clinical Product Advisor / Specialist, and Clinical Workflow & Implementation Advisor engagements (Remote Global US/EU/UK Contractor & EU Relocation).