Full-stack engineer in San Diego, 15+ years building for the web. Founded RipeMetrics in 2017 and spent 8 years leading product and AI engineering on a marketing automation, business intelligence, and customer support platform — which is where I learned how to ship AI features in production and what it actually takes to keep a SaaS product alive long term.
I'm currently going deep on agentic engineering — Claude Code, autonomous workflows, human-in-the-loop patterns, and the infrastructure that makes AI systems reliable rather than just impressive in a demo. This profile is the working surface for that: the repos below are the actual systems I use, organized by how they relate.
Agentic engineering: Claude Code · MCP · GitHub Copilot · autonomous loops · skills & instruction layers · LiteLLM gateway · sandboxed execution
AI & LLMs: OpenAI · Anthropic · Azure OpenAI · RAG · embeddings · vector search · tool calling · structured output (Zod) · prompt engineering · LangChain · Pinecone · Chroma · pgvector · fal.ai · Ollama
Frontend: TypeScript · React · Next.js · Astro · Redux · RTK Query · Tailwind · shadcn/ui · Zod
Backend: Node.js · Python · FastAPI · Express · Laravel · PHP · Koa · tRPC
Data: PostgreSQL (+pgvector) · MySQL · Redis · SQLite · Drizzle · Cloudflare D1 · Supabase
Cloud & DevOps: Azure (OpenAI · Blob Storage · Functions) · Cloudflare (Workers · D1 · R2 · KV) · AWS (in progress) · Docker · GitHub Actions · CI/CD · Vercel · Railway
Integrations: GoHighLevel · WooCommerce · Stripe · Authorize.net · PayPal · Twilio · ElevenLabs · Shopify · ShipStation · Sanity · Resend · SendGrid · PostHog · Sentry
Currently studying: Azure AI Foundry · OpenAI Agents SDK · Google ADK · Vertex AI · AWS Solutions Architect Associate
IndyDevDan's agentic engineering track — Tactical Agentic Coding, Principled AI Coding, Agentic Horizon · Claude Code for Real Engineers (Matt Pocock) · Model Context Protocol (fundamentals · advanced · server UI · auth) · AI Engineering Bootcamp (Maven) · LLM Engineering: Foundations to SLMs (AI Makerspace) · Cursor AI Bootcamp · Frontend Masters (React · Next.js · TypeScript · Redux) · AWS Solutions Architect Associate (in progress)
The core of what I'm working on: infrastructure that makes AI coding agents reproducible, context-aware, and safe across every machine.
- ctrl+shft — the system your AI agents are missing: synced instructions, workflow skills, hardened secrets, and an autonomous loop that picks GitHub issues, implements them, and commits while you're AFK. Planning pipeline (grill-me → write-a-prd → prd-to-issues → do-work) and a compliance HUD included.
- llm-gateway — route Claude Code through a GitHub Copilot subscription via a secure LiteLLM proxy.
- cmd — a markdown-first business operating system for solo operators and AI agents; strategy and clients in plain files an agent can act on.
- ctrlshft-hub — the single source of truth for the ctrl+shft agent layer — composite actions, reusable workflows, and the TypeScript engine that bootstraps consumer repos (evolved from sandcastle).
Learning agentic coding in public, one module at a time.
- agentic-nlq — deterministic code orchestrating non-deterministic agents: an agentic layer (specs,
.claude/commands, ADWs) around a natural-language-to-SQL app. The Tactical Agentic Coding arc subject. - Tactical-Agentic-Coding — the official Agentic Coding course by @IndyDevDan, mirrored for study; progress on arndvs.com/blog.
A fictional smart-bed company built as a sandbox for AI video generation via agentic workflows — short film, product teaser, and a working website.
- PUSH — PUSH — a short film. One morning. One button. Push Mode cannot be manually interrupted once initiated. This is a feature, not a limitation.
- riseawake.com — the website for RISE Technologies, Inc. — makers of the RISE Push.
- rise-co (private) — ongoing RISE company work.
- cast — creative automation pipeline for generating localized social ad creatives at scale.
- cast-backend — backend pair for cast.
- reel-decoder — decodes Instagram reels into structured Swipe Library rows locally (Ollama, zero cloud APIs).
Website for a San Diego chiropractic clinic, built well beyond what the brief required. Next.js 16, Sanity CMS, 2,000+ TypeScript files. Five AI integrations: GPT-4o Vision insurance verification, OpenAI content enhancement pipeline, LLM review sentiment analysis, AI spam detection, and AEO endpoints. Custom JSON-LD composition architecture with 76 Schema.org types including medical schemas with ICD-10/SNOMED-CT codes. Programmatic SEO engine generating 203+ pages. 55 automated tests, 4 CI/CD pipelines.
Active client engineering — sales site, e-commerce, content operations, and agent infrastructure. Systems-with-guardrails: an AI editorial pipeline where "approved" is a SHA-256 fingerprint of copy + assets, a fact-check pass that caught 8 fabricated citations before publication, a human-in-the-loop reply assistant where every send is operator-approved, and an order pipeline rebuilt around "the database is the record of truth" after an incident shipped ~$381 of product to people who never paid. From-scratch quality baseline (storefront went from no tests to 195) and a 284-item audit of things that "looked worked but didn't" — each closed with a fix or a written decision. The backlog runs on a label-driven agent pipeline through my own LiteLLM proxy.
- mcrdse-site — the main sales site
- mcrdse-super-market — e-commerce / storefront
- MCRDSE-Content-Ship — content production + shipping pipeline
- mcrdse-outreach — outreach automation
- mcrdsemovement-site — movement/community site
- mcrdse-command — command-layer ops
- mcrdse-ops — operations infrastructure
- hermes-ops — agent ops
- launch (private) — mobile-first monorepo starter I fork into client products
- foreword (private) — idea to development-ready spec
- scorpion (private) — Scorpion Percussion app
- aligned (private) — healthcare-aligned products
Private client codebases; the pattern is public in principle.
Small, focused things I built because I needed them.
- tailwind-indicator — zero-dep screen-size indicator for Tailwind CSS (breakpoint, dimensions, orientation, pixel ratio).
- preact-chatbot — multi-tenant AI chatbot widget — Preact island, Shadow DOM isolation, Pusher streaming.
- comedian-voices — agent skills for writing in famous stand-up voices — structural moves, rhetorical patterns.
- voiceprint — voice style generator — turns raw material into loadable SKILL.md writing skills.
- kitchen-skylight — family kitchen skylight display — calendar, recipes, home automation hub (OpenSkyLight-based).
- dynacraft (private) — dynamically craft customized CVs and cover letters for job applications.
- kenesis-360-keyboard-layout — alternative layouts for a split ergonomic keyboard.
- arndvs-py-api — generic Python API backend — FastAPI on Railway, powering arndvs.com services.
Learning projects and older experiments (mostly historical — kept for reference): the AI Engineering Bootcamp repo, LangChain/Chroma RAG experiments, React/Redux learning repos, Laravel courses, and assorted early web work. The current direction is the agentic engineering stack and applied AI above.
Notable from the RipeMetrics years (2017–2025, now wrapped): tenant-isolated RAG with OpenAI/Anthropic provider fallback so one vendor going down never took it offline, an AI onboarding pipeline that cut client setup from ~30 minutes to under 60 seconds, and a Laravel → React/Next.js migration (880 components, 85 routes) done behind feature flags without downtime.
- Building for the web since 2010 — HTML/CSS → Joomla → WordPress → Laravel → React/Next.js
- Founded RipeMetrics (2017–2025) — led product, frontend, and AI engineering across a team of up to 12
- Based in San Diego, open to remote
arndvs.com · linkedin.com/in/arndvs
These are personal projects. The views and work here are my own and unrelated to my past, present, and future employers.
Private repos are marked (private) — they're client or personal code not shown publicly.