§ 02

AI Work

Shipped,
not specced.

Most of what’s here runs on production traffic or was built and deployed end-to-end — formal AI training put to work, not left on a certificate. AI fluency is increasingly table stakes for senior product work, and this is what that looks like in practice.

§ 02.1

Production · Exclusive Resorts

AI shipped at scale
in a luxury travel platform.

Some member-facing, some internal tools for advisor teams — production releases and active betas. The member-facing tools draw on CMS and concierge-vetted content first, before broader sources.

What we’ve built

Trip Recommendation Tool Members describe what they’re after; the tool surfaces matched destinations and itineraries from live inventory.

Trip Itinerary Creator Builds full activity and dining itineraries from known member preferences and a few quick questions.

Member Profile Summarization Profile summaries giving advisor teams instant context on preferences and travel history — less ramp time before every interaction.

AI Chat Agent Conversational agent handling member inquiries with access to property data, availability, and member records

LLMsRecommendation SystemsRAGAWS Bedrock

§ 02.2

Personal Project · Built in Bolt

CrewChief — AI-powered
automotive ownership platform.

Full-stack web application I built to solve a real problem: most vehicle owners have no organized system for their car’s knowledge, history, and future work. CrewChief fixes that with a persistent AI advisor that actually knows your car.

Core concept

When you add a vehicle, the app calls Gemini to generate a comprehensive dossier — known failure points with mileage ranges, full factory maintenance schedule, fluid specs, common modifications, reliability score, and powertrain details. Everything downstream builds from that knowledge base.

crewchief · My Garage

Under the hood

AI Consultant

Gemini 2.5 Pro chat initialized with a 200+ line system prompt injecting the full vehicle record — mileage, service history, tracked issues, mods, and recalls.

Agentic Write-Back

Structured command tags emitted mid-conversation, parsed and executed server-side — adds wishlist items, closes issues, updates mod status without leaving chat.

Model Tiering

2.5 Flash (structured JSON, temp 0.3) · 2.5 Pro (conversation, temp 0.7) · 2.0 Flash (vision/OCR) — allocated by cost and capability.

Invoice OCR & History

Upload a service invoice; Gemini 2.0 Flash extracts vendor, date, mileage, and every line item into timestamped records the AI references in future sessions.

Cost Estimation

Itemized parts + labor ranges with regional rate acknowledgment, validated against a strict Zod schema — no hallucinated line items.

Labor Bundling Optimizer

Items tagged by physical zone; bundler flags shared access requirements and estimates savings from co-scheduling.

Next.js 13TypeScriptSupabaseGemini API

Source → Live app →

§ 02.3

POC · Exclusive Resorts

Lumière Collective — a POC built to test
onboarding drop-off.

We were seeing member drop-off in the community onboarding flow, right at the profile photo step. Working build, demoed internally, not yet rolled out to members.

The hypothesis

The bet: if a member doesn’t have a photo they’re confident in, they abandon and don’t come back. This POC gives them an exit — upload or capture a portrait, get four professionally styled variations in under 60 seconds.

Under the hood

Parallel AI Generation

Four simultaneous Gemini 2.5 Flash calls via Promise.all() — Natural, Corporate, Editorial, Vacation — results returned together, ~75% faster than sequential.

Edge Function Proxy

Deno runtime Edge Functions: API key management, server-side safety validation with keyword/negation parsing, automatic person-removal pre-processing, CORS.

Dual Safety Layer

Client validation blocks immediately; server-side Edge Function pass catches any circumvention via keyword/negation context parsing.

Camera & Upload

MediaDevices API live capture, drag-and-drop upload, before/after slider comparison, one-click download. HEIC/PNG/JPEG support.

Custom Luxury UI

No UI library — hand-authored Tailwind + CSS keyframes: magnetic cursor on rAF, Ken Burns hover, glassmorphism toast, cinematic loading sequence.

Image Pipeline

Client-side resize before upload; base64 encoding in chunked 8 KB passes to avoid call stack overflow on large files.

Lumière Collective · Atelier

Lumière Collective — The Portrait Atelier

React 18TypeScriptTailwind CSSGemini API

Source → Live app →

§ 02.4

AI Practice

AI as a working environment,
not a feature.

This is harder to show than a shipped product, but it’s probably more relevant to how I’d operate on your team.

Claude Teams — how I actually use it at work

I use Claude Teams across a significant portion of my work, and I’ve built out a structured practice around it — not just prompting, but building shared tooling for my team. The practical effect: tickets, release notes, and interview synthesis start from a structured first pass in minutes instead of a blank page — and the whole team drafts from the same structure.

Projects per product A dedicated Claude project for every product I directly PM — each with robust system instructions, role context, and shared skills so teammates can pick up context without a handoff meeting.

Workflow skills Custom skills for Jira ticket writing, release notes, and standup prep — shared across the team so everyone drafts from the same structure and tone.

Hiring tooling Projects with skills for synthesizing interview notes and evaluating candidates against role requirements — structured output, consistent format across the panel.

Scheduled jobs Automated recurring tasks — weekly digests, status rollups, pre-review prep — running on schedule without manual setup each time.

MCP integrations Connected to Jira, Confluence, and Figma via MCPs — writes tickets, drafts documentation, and captures design notes directly from conversation.

Cowork projects Cowork sessions scoped to each product I own — context, artifacts, and ongoing work organized in one place instead of scattered across tabs and docs.

Claude TeamsMCPPrompt EngineeringLLM Workflows

Shipped,not specced.

AI shipped at scalein a luxury travel platform.

What we’ve built

CrewChief — AI-poweredautomotive ownership platform.

Under the hood

Lumière Collective — a POC built to testonboarding drop-off.

Under the hood

AI as a working environment,not a feature.

Claude Teams — how I actually use it at work

Shipped,
not specced.

AI shipped at scale
in a luxury travel platform.

CrewChief — AI-powered
automotive ownership platform.

Lumière Collective — a POC built to test
onboarding drop-off.

AI as a working environment,
not a feature.