Case Studies
Each of the five projects on my portfolio, laid out with the video, what it does, the stack, and the honest note on how it got built. This is written in my voice so you can hear it the way I would say it to you in a call.
The private AI console that runs the business side of my channel so I can spend my time making videos.
I built Daryl as my personal operations manager for the YouTube side of my work. It runs on a Cloudflare Worker with an agentic tool loop that talks to the Claude API. It has persistent memory through Supabase and Voyage embeddings, so it learns me over time and remembers what we talked about across separate sessions.
Daryl pulls my live channel stats through a Google OAuth flow I set up end to end. It reads my Gmail and finds sponsorship outreach I have missed sitting in spam or promotions. It drafts pay-first replies in my own voice and holds them in Gmail drafts for me to review. It never sends on its own. Sending is the one line I keep on my tap, on purpose.
The judgment about what Daryl is, what it should do, who it should serve, and where the lines are, is mine. I wrote the dossier that defines its mission and its guardrails. I gathered every API key, ran the OAuth Playground flow myself, executed the Supabase pgvector SQL, and debugged the whole thing live when the memory round-trip and OAuth tokens misbehaved. The Worker code was written with AI to my direction, and I own the architecture, the reviews, and the decisions.
The direction, the mission, the integrations, the OAuth, and every deploy click is mine. The heavy Worker code was written with AI to my direction, and I own the architecture and the reviews.
Daryl itself stays private because it is a personal tool. The demo above and the dated build record below are the public windows into it.
A subscription consumer product I took from a raw idea into a live, deployed waitlist on its own domain.
GasElite is a subscription consumer product I built from the initial concept into a running product on its own domain at gaselite.online. The waitlist itself is live and already collects real signups. The full product is in stealth, so I do not describe its features here on purpose.
Everything a small team would ship in a quarter is on that site. Email and phone capture with strict client and server validation across the US and 132 international phone formats. Automated branded confirmation email through Resend. A contact form with validation, a honeypot, rate limiting, and an admin dashboard. Privacy, Terms, Contact, and Unsubscribe pages with real compliance detail. Full brand assets including a 3D interactive hero logo, home-screen icons, and an OG image.
The infrastructure work is the part I am proudest of because I did every click of it myself. I bought the domain, moved DNS to Cloudflare, set up the Redirect Rule that got the apex loading in about half a second, verified the sending domain in Resend with DKIM, SPF, and MX records all green, added the DMARC record, turned on Cloudflare Email Routing so hello@ receives and forwards, ran the deliverability tests, and rotated every exposed key. The waitlist ships to real people right now.
The product idea, the data-sourcing strategy, the pricing, the rewards system, the affordability tiers, the legal decisions, and every deploy click is mine. I wrote a 13-page data-sourcing strategy myself as the ground truth for the whole data layer. The application code was written with AI to my direction.
The waitlist is public. The full product itself is in stealth and I do not describe its features on public documents on purpose.
A personal daily protocol tracker I built for my own use, shipped as an offline Progressive Web App.
The Hair Recovery Command Center is a personal daily protocol tracker I built for myself. It ships as a single self-contained HTML file that runs offline, installs to a phone home screen through the Web App Manifest, and stores every check-in locally. Nothing leaves the device. Nothing needs a network to keep working after the first load.
I built this the way I wish more personal tools were built. The visual identity is deliberate, the whole thing is one file, and it respects the routine it is tracking. Today view for what I need to check off. 28-day calendar for the bigger picture. Weekly plan view for the core protocol. Progress check-ins every four weeks with shedding and irritation scoring, notes, and a photo prompt. A safety dashboard for the clinician-guided rules that matter. Program settings so I can set my own start date, pill time, and microneedling day.
If you want to see how I ship polished consumer tools in vanilla web tech with zero framework overhead, this is the clearest example on the portfolio.
This is a personal tool for my own use, not a medical device and not medical advice. Prescription labels and clinician instructions always override the planner. I built it because I wanted a tool that respected my routine and worked offline.
The design, the UX flow, the data model, and the PWA engineering are mine. I own it end to end.
A curated tool I built to make my own remote job search organized and honest about its sources.
RemoteFit is a curated tool I built to make my own remote job search organized and honest. I defined the fit logic myself, categorized 100 roles across categories, tagged résumé variants against each so I know which version to send where, and built application-status tracking with direct employer-search handoffs. A scattered browser-tab job hunt turns into one ranked, trackable view.
I built it because I was actually running this search. The choices about categories, ranking, and tagging are the parts that decide whether the tool is any good, and they are all mine. The tool does not scrape and it does not auto-apply. The list is manually curated on purpose so I know exactly what is on it and why.
The role list is manually curated, not a live scraped feed. Individual job links can expire and should be verified before applying. This is a working prototype I built for my own use during my job search.
The fit logic, the categorization, the tagging system, and the UX are mine. The application code was written with AI to my direction.
My personal Python learning series, where I show my hands-on code and the honest arc of learning it while building.
F800 is my personal Python learning series. Two models, both command-line diagnostics for capability scoring and system-health calculations. Model 001 integrates Gemini for the analysis step. Model 002 documents the iterative Python learning behind the series. This is where I show my hands-on Python and the honest arc of learning it while building.
The two videos are side by side on purpose. The story here is the progression from Model 001 to Model 002, and the value is the practice, the iteration, and the honest write-up of what I learned along the way.
This is a learning project, not an autonomous AI system and not an enterprise product. The Python is personally written. The AI was integrated by me. The value here is the practice, the iteration, and the honest write-up.
Model 001 is on my GitHub for anyone who wants to read the code.