Electronic Health Record for Independent Practices

Monolithic EHR designed for 1-5 provider private practices. Real InterSystems IRIS Native persistence (not a Mongo-shaped mock), Express + TypeScript backend, React + Vite frontend, Lerna monorepo. Modules wired end-to-end against IRIS: auth, patients, scheduling, clinical notes/vitals/allergies/problems, audit, billing (charge / claim / payment with state machine), inbox, medications. FHIR R4 compliance layer and Bulk-FHIR interop in progress.

TypeScript InterSystems IRIS Native Express React 19 Vite 7 Tailwind 3 Lerna

• Real IRIS Native driver via @intersystems/intersystems-iris-native (extracted at build time, not committed)
• Audit logging at the repository layer - every PHI read/write captured to a dedicated global
• Custom design system (heme) - three-anchor palette, no semantic colors, no pill shapes, status by icon + label + weight
• CI smoke job spins an IRIS container, vendors the driver, validates round-trip persistence on every PR

Unified Disease-Knowledge System

Keys narrative clinical reference content to a canonical medical code spine, so a diagnosis, medication, or lab code resolves to ranked pathophysiology, diagnosis, and treatment knowledge with crosswalks across every major coding system. The internal key is the UMLS CUI; ICD-10-CM, ICD-11, SNOMED, RxNorm, and LOINC are stored as projections, reconciled with Mondo for disease-class precision. Surfaced through a FHIR R4 terminology service and EHR-friendly delivery, with a citation-faithful GraphRAG layer built last. The release gate is citation fidelity: no clinical claim ships unless it traces to a verifiable source span.

Python TypeScript PostgreSQL pgvector Apache AGE HAPI FHIR Mondo / SSSOM Claude Citations

• UMLS-CUI spine with crosswalk projections across ICD-10-CM, ICD-11, SNOMED, RxNorm, LOINC, Mondo, HPO
• FHIR R4 terminology service ($lookup / $translate / $expand / $validate-code) on HAPI
• Open-core / hosted-premium split enforced by a CI redistribution guard and runtime row-level entitlement
• Citation-fidelity acceptance gate: every generated clinical claim carries an entailment-tested citation

Longevity Intelligence Platform

Multi-agent AI system for longevity optimization targeting insurers to demonstrate reduced claims via wearables.

TypeScript Python Multi-Agent Orchestration Real-time ETL

• 400+ biomarker integration
• 30+ concurrent AI agents
• Multi-omics data pipelines
• HIPAA-compliant infrastructure

Autonomous AI Research Loops for Web Applications

Adapts karpathy/autoresearch for webapp optimization. An AI agent autonomously modifies your code, runs evaluation, keeps improvements, reverts failures, and repeats — overnight. Published on npm.

TypeScript EJS Node.js CLI Lighthouse Vitest ESLint

• 9 research modes: perf, quality, security, a11y, clean, wiring, styling, marketing, feature
• Auto-detects framework (Next.js, Nuxt, SvelteKit, Vite, Remix)
• Auto-detects integrations (Supabase, Prisma, Stripe, Clerk, shadcn)
• Configurable guardrails with HIPAA compliance support
• ~60-100 autonomous experiments per overnight session

-> View Repository · npm

AI-Powered Pre-Med Guidance System

Multi-persona AI mentor synthesizing wisdom from nine distinct advisor perspectives into personalized counsel for pre-medical students.

TypeScript Next.js 14 FastAPI Claude API PostgreSQL pgvector

• 9 advisor personas with dynamic weighting
• Multi-persona synthesis engine
• Opportunity matching with vector similarity
• Commitment tracking and accountability

-> View Repository

Design System Extraction Plugin

Claude Code plugin that extracts design systems from live websites using Playwright, outputs structured JSON, and generates Tailwind configs, CSS variables, shadcn themes, and W3C tokens.

Node.js Playwright CIE ΔE 2000 WCAG Auditing

• Live URL extraction via browser automation
• 8 pre-extracted references (Stripe, Notion, Linear, YouTube)
• Multi-format output (Tailwind, CSS vars, shadcn, W3C, Figma)
• Project comparison with gap analysis

-> View Repository

Medical Provider Certification Platform

Comprehensive resource for medical providers seeking peptide therapy information, certification, and verified supplier connections.

Next.js TypeScript Supabase Tailwind CSS

• Searchable peptide database
• CME-accredited certification courses
• Verified supplier directory
• HIPAA-compliant professional forums

-> View Repository

Real-time 3D Multi-Entity Visualization

High-performance monitoring system rendering 100+ entities at 60fps with near-zero latency control.

Three.js React Three Fiber WebGL Instancing Web Workers WebSockets Redux Next.js 15

• 1000+ updates/second throughput
• Worker thread spatial indexing
• LOD + frustum culling optimization
• Protocol Buffers serialization

-> View Repository

Universal Code Indexer for AI Assistants

Enterprise-grade codebase indexer with parallel processing. Published on npm.

TypeScript Worker Threads Tree-sitter AST Parsing Claude SDK

• 3.6x faster with Worker Threads
• 10,000 files in ~25 seconds
• 40% memory reduction
• 9 languages supported

-> View Repository · npm

MCP Server for Spa/Wellness Management

Model Context Protocol server enabling AI agents to manage Zenoti operations - appointments, guests, billing.

TypeScript MCP Protocol REST API Zenoti API

• 4-step service booking in single call
• Guest lifecycle management
• Invoice & payment processing
• Agent-optimized responses

-> View Repository · npm

Medical Practice Platform

Omnichannel patient communication system with practice management integration.

Next.js Twilio Zenoti Supabase HIPAA

• Facebook Messenger
• Google Business Messages
• SMS aggregation
• Zenoti booking sync

Marine Wind Prediction

Real-time wind sensor and prediction model positioned 100m from shore.

Python Time Series IoT Weather APIs

• Sailing enthusiasts
• Marine businesses
• Race committees
• Harbor operations

-> View Repository · Live

Open Verification Framework for AI-Generated Claims

A fail-closed grounding layer: a generated claim may act only when it carries a warrant the gate can re-check against the domain's source of truth, and the system refuses by default otherwise. Distinguishes weak referential warrants (a pointer to a source) from strong derivational warrants (re-running the underlying query, test, or computation), and routes the important-but-unverifiable case to a human with the evidence laid out. Principles, synthesis, and architecture are published in the repository; two companion preprints are on Zenodo. The first vertical, SMS delivery-truth and response attribution, is in implementation. Built vertical-first, the portable contract is extracted only once two real verticals share it.

TypeScript

• Fail-closed gate: refuses by default unless a claim carries a re-checkable warrant
• Referential vs derivational warrants - pointer to a source vs re-running the query/test/computation
• Important-but-unverifiable claims routed to a human with evidence laid out
• Two preprints: Litmus: The Admissible Set · System Architecture
• Vertical-first; portable contract extracted only once two verticals share it

-> View Repository