I'm a performance marketer who learned to ship software
by building real products end-to-end.

The feature I'm most proud of is the SOS craving flow. When a craving hits, Phase 1 grounds the user with breathing (mandatory for high-intensity cravings, because biology before morality). Phase 2 opens the toolbox, breathing, cold water, distraction, reasons, or talking to Milo. Phase 3 logs the outcome. A slip is data, not failure. That decision comes straight from CBT and Motivational Interviewing, and it runs through every line of copy in the app.

The bigger system behind Milo is the content pipeline. Sonnet writes drafts, Opus runs the quality gates, clinical verification, citation checks, voice consistency, 10 Sanity schema rules, the whole thing. Fifty-one articles in two weeks, every one passing the clinical gate. Not "ChatGPT writes a blog post." A real production pipeline with model routing discipline.

I mapped the competitive landscape on two axes: nicotine-specific vs. generic, and active AI vs. passive. Sunflower has 500K users but zero nicotine focus. Kwit has 14 years of content but zero AI. Smoke Free is a tracker. Milo sits alone in nicotine-specific + active AI. The content strategy targets the white space Sunflower left wide open.

The architectural decision I'm proudest of: the website and the iOS app share the exact same Supabase backend. Same tables, same Edge Functions, same embedding index. Schema changes ripple to both platforms by design. One source of truth, no data divergence.

The closed loop is the best part. The iOS app has a tab called CoCreate: users brainstorm yoga episode ideas with an AI teacher. The teacher detects when an idea has substance, surfaces a "Submit to Queue" button, and writes directly into the content_queue table that feeds the podcast pipeline. Scout picks the idea up, Curator approves it, Scriptwriter generates a source doc, Artist creates cover art, I upload to NotebookLM, publish to Spotify. Distribution clips feed Reddit and YouTube, performance metrics feed back into Scout's topic scoring. Users brainstorm, the podcast produces, distribution promotes, performance informs the next batch. The app feeds its own content engine.

The semantic search is a single 768-dimension HNSW index on pgvector, queried by both the Next.js site and the iOS app via the same Edge Function. Hybrid scoring with Reciprocal Rank Fusion combines vector similarity and full-text relevance. "Ancient wisdom for modern stress" returns the right episodes, not because the keywords match, but because the meaning does.

This is the workspace I'm most proud of, because it's the system that makes everything else shippable by one person. The core innovation is a three-layer instruction hierarchy: a root file that defines how I work with the AI, a workspace file that maps every project and how they relate, and per-project files with the technical specifics. Every session starts with the right context. Zero ramp-up, no repeated explanations, no drift between conversations.

The second piece is model routing discipline. Writing and procedural work runs on the faster model, maybe eighty-five percent of tokens. Judgment calls, verification, clinical review, and quality gates run on the stronger model. Five to fifteen percent of spend, but that's where the cost returns live. This isn't a theoretical optimization. It's how I keep one-person AI infrastructure economically viable.

The third piece is the RUNBOOK pattern, per-project operational files that encode exactly how to deploy, where credentials live, known failure modes, and what "done" means for each task type. Instead of tribal knowledge in my head, it's documented procedure the agent can follow. If I'm unavailable and someone else needs to operate the system, the runbooks are the manual.

The workspace supports five active project areas running in parallel: two iOS apps in production, two websites, two content pipelines, competitive intelligence, and this portfolio. Same agent, same standards, different contexts. This is what AI-native development actually looks like when you build it on purpose.