Peter Dirickson

Identity

I'm Peter Dirickson. Brazilian-American, based in San Diego, with one foot in São Paulo and another on the Península de Maraú in Bahia. Senior Enterprise Architect at Scripps Health by day. Founder of Frenti, a one-person AI-native product agency, by night and weekend. I've been writing software since 1986 — twenty-five years professionally, the last four with Machine Learning and AI.

Now

Building BonVivant — a San Diego lifestyle and surf guide that blends restaurants, neighborhoods, and activities into one editorial layer.

Building VisitMaraú — a marketplace for the Península de Maraú in Bahia, Brazil, where I also run a beach house called Zoetry.

Path

Started writing software in São Paulo in 1986 on a TK 85 — Microdigital's Brazilian clone of the Sinclair ZX81. Healthcare and enterprise systems through the 90s and 2000s. Moved to San Diego in 2004. Senior Lead Software Engineer at Hewlett-Packard from July 2010 to August 2015. Senior Enterprise Architect at Scripps Health since October 2019. Founded Frenti in 2024 to ship AI-native products on the side.

Scripps Health

I'm a Senior Enterprise Architect at Scripps Health, one of San Diego's largest health networks, since October 2019. The work is architect-first: I design and ship enterprise systems on a Ruby on Rails core with Azure infrastructure, integrating across legacy healthcare platforms. I also build — recent work includes a Playwright-based visual regression monitor and a handful of internal tools.

I led the introduction of agentic coding to the engineering team. Stood up Anthropic models via Azure AI Foundry and established how agentic patterns get applied inside enterprise healthcare constraints — security, compliance, change management.

Software House of One

Frenti is a one-person agency with agentic coding and a spec-driven development discipline that lets me ship like a five-person team. Most of my products are mine; I take a small number of client engagements per quarter.

Skills and craft

Tech

• TypeScript and Next.js (App Router). Default stack for every Frenti product. Server Components by default, Client Components only where animation or state requires it.
• Tailwind CSS v4. Design tokens in @theme, no tailwind.config.ts. Atlantic Dusk palette across the personal site.
• React + Framer Motion. Single animation library per project — no mixing CSS keyframes with JS animation.
• Python. Used for editorial pipelines, data ingestion, and ML workflows. Twenty-five years of professional software, the last four with Machine Learning and AI.
• Postgres + pgvector. Supabase for managed Postgres, RLS for row-level multi-tenancy, pgvector with HNSW index for RAG retrieval.
• Vercel. Fluid Compute for the API, AI Gateway for model routing, Edge Config and Blob where they fit. No Edge Functions — Node.js runtime is the default.
• AI SDK + AI Gateway. streamText, embed, generateObject. Provider-agnostic model strings (anthropic/claude-sonnet-4-6, openai/text-embedding-3-small) routed through the Gateway.
• MDX content collections. Long-form content lives in MDX with validated frontmatter. Same file feeds the rendered page and the RAG index.
• Zustand + localStorage. Client state when Server Components can't carry the load. Persisted under a versioned key.
• Ableton Live + Max for Live. Side B — deep house production.

Craft

• Spec-driven development. Every non-trivial piece of work starts as a markdown spec under specs/ or docs/plans/. Goal, phases, acceptance criteria, code references. The spec is the contract; the implementation follows.
• Agentic coding with Claude Code. Most of my output is Claude-paired. The leverage isn't in the typing — it's in the spec, the review, and the judgment about what to ship.
• LLM-as-judge editorial gates. Generated content (restaurant copy, neighborhood voice, itineraries) is scored by a separate model against tasting notes — factual accuracy, voice match, no hype words. Anything below threshold goes to manual review.
• Voice profile per domain. Each product has its own voice.md (and sometimes one per neighborhood, like in BonVivant). Voice files are style only — identity facts live in the indexed content so they can be cited.
• RAG with grounding rules. Chat answers are constrained to retrieved chunks. The model is instructed to acknowledge gaps rather than invent. Same pattern across personal site and product chatbots.
• Bilingual systems. English and Brazilian Portuguese first-class, not bolted on. i18n routing, per-locale MDX, locale-aware retrieval.
• Editorial pipelines. Multi-phase Claude Code pipelines that take structured data in, run it through voice-aware enrichment and judge scoring, and emit publication-ready MDX.

Side B

I produce deep house in Ableton with Max for Live, late at night when the house is quiet. I'm a Rotary Club member in San Diego. I travel between San Diego, São Paulo, and Bahia. A local host manages Zoetry on the ground while I run it remotely.