← PREV {{ prevProjectTitle }}

NEXT {{ nextProjectTitle }} →

The first version of Baroline had already shipped. Drivers used it for a few weeks, then stopped. Went back to phone calls.

Inheriting a failed product is different from starting from scratch — the constraint isn't technical, it's trust. Trust, once broken, doesn't come back because the screens got cleaner.

Highway. Night. Job update comes in. Pull over. App loads on bad signal. Three menus. Confirm a status. Screen times out. Signal drops. Nothing saves. Your operator calls because the system shows nothing.

"I'm not crashing my truck for this."

First week: full team, every feature from the old system written on cards. One question per card: driver or operator? Sixty cards. Six landed in the driver pile.

Before that sort, I ran the full feature list through an LLM — pulled every function, tagged each as driver-facing or operator-facing, stress-tested my own logic against it. The cards were the physical version of a call I'd already pressure-tested. Faster to argue a brief when you've already checked your reasoning against something that doesn't care about your ego.

Accept job. Start route. Arrive. Call operator. Problem report. Complete delivery. Everything else was operator complexity that had shifted into the driver app. Nobody had forced the question of who was actually paying for it.

PM — "Drivers need metrics—"

ME — "We're designing for people who actively don't want to use our product. Let's make it so simple they barely have to."

PM — "So we're designing an app for people who hate apps?"

ME — "Exactly."

Drop image, video, or GIF

profile screen

OLD — PROFILE

Drop image, video, or GIF

loads screen

OLD — LOADS

Drop image, video, or GIF

dashboard

NEW — DASHBOARD

Drop image, video, or GIF

loads detail

NEW — LOADS

OPERATOR — "This is just my notebook."

ME — "Yeah. Because your notebook worked."

Drop image, video, or GIF

operator's spiral notebooks

Drop image, video, or GIF

new digital workflow

Felt obvious. In testing, two out of three operators moved jobs to the wrong day — a truck arriving at a factory that wasn't expecting it.

I rebuilt it: select the job, choose the driver, confirm with a plain summary. More steps. Zero mistakes in the next round. Intuitive and safe are not the same thing. In a logistics system, safe wins every time.

The driver flow worked. Cut from 15 functions to 6, drivers completed the core flow first try — the thing the old app never got. Then I left before full launch. The product failed anyway. Not the design. The product. These are different things. Products are systems — business decisions, funding, timing, internal politics, things a designer never fully sees. Good design is one variable. Not the only one.

What I carry: complexity never disappears. You decide who carries it. Never let it be the person doing their actual job.

Before taking my proposal to the team, I ran my logic through Claude and Grok—not for validation, but for attack.

I laid out the drop-off metrics, my proposed pivot, and asked both models to tear my direction apart. I wanted to expose every weak spot before speaking up in a room full of people who had been working on this far longer than I had.

Saying "stop building this" in week three, as the person brought in to lead, felt like admitting I was lost. I said it anyway.

Drop image, video, or GIF

old, static roadmap page

Drop image, video, or GIF

Azure workspace / infrastructure screenshot

Shifting the backend to Azure solved the tech constraints, but the real fight was over the funnel interface. The team wanted maximum speed—fewer inputs, frictionless signups.

I pushed back.

Our data proved users weren't lacking resources; they were drowning in them. They didn't need another fast, generic output—they needed to feel understood.
I fought for intentional friction. I believed that if we asked the right contextual questions and got the user to invest a few moments of effort, our AI could deliver something truly life-changing.

Drop image, video, or GIF

user survey — friction research

Drop image, video, or GIF

old funnel landing page

Figma was slowing our iteration cycle down. Not because of the software, but because of the review loop. In a small team, we still couldn't communicate interaction feel without three rounds of "Wait, how does this actually work in practice?"

I had been following bolt.new, an AI coding tool that lets you prompt full-stack interfaces into existence. I proposed building directly in code first to see what broke, using Figma later only to align with engineering once the mechanics were proven.

I introduced it casually, but I needed it to work. It did. We had a functional interface in hours. The team moved from reviewing static frames to testing live interactions—we stopped describing ideas and started living in them.

Drop image, video, or GIF

Midshift landing page build

Because our intentional friction gathered deeper context, the AI could finally do real work. Instead of crushing users under a skill-dense PDF, we decided to show them where they actually stood in the global market.

We integrated live market data APIs to cross-reference their extracted skills against real-time demands across the US, Canada, and the UK. The interface immediately surfaced their competency level, salary benchmarks, and market fit—proving the concrete value of our analysis before asking them to create an account.

This value-first moment built the necessary trust. By the time the interface guided them to sign up for full AI-powered career coaching, it didn't feel like a barrier; it felt like the logical next step for their career.

Drop image, video, or GIF

competency / market fit surface

Drop image, video, or GIF

salary benchmark view

Drop image, video, or GIF

sign-up moment, full view

None of this was a stroke of genius—it was dozens of small corrections driven by tight feedback loops. It was messy, fast, and demanding.

I learned something critical about AI products: user trust doesn't come from an LLM being "smart" or impressive. It comes from highly curated outputs that mapped directly to an individual's skills.

By delivering clear, immediate actions instead of vague career advice, we didn't just ship a smart tool—we built a product that respected the user's time and their career pivot.

Drop image, video, or GIF

Admin working panel — real-time modular block customization

2×

My time at Bettermode proved that high velocity doesn't mean cutting corners—it means building personal leverage. By taking full ownership, embracing AI tools as research partners, and respecting system constraints, I delivered enterprise-grade work at startup speed without needing to be managed.

Turning design audits into business ROI.

Teardown isn't just an analyzer; it's a personalized audit tool that trains itself on your specific business goals. Built builder-first — a sophisticated AI stack that moved this from idea to functional web app at record speed.

Drop a video, GIF, or image

Cinematic hero — Teardown product demo (or click to browse)

From idea to deployable app, one AI stack at a time.

Discovery-first, not checklist-first.

To transform raw audits into actionable intelligence, I developed a system that contextualizes design decisions through the lens of user intent. This engine serves as the backbone for the diagrams and data-driven insights detailed below, ensuring that every UI recommendation is backed by structural logic and business impact.

Drop a video or GIF

Deep context scanning & annotation system in action (or click to browse)

UIKit: decoding brand DNA in a single click.

We've all been there — joining a new team or starting a redesign only to spend the first week on a scavenger hunt through Figma, Google Drive, and old Slack threads just to find the right logo or hex code. It's a massive time sink. UIKit solves this by using AI to instantly extract every component, token, shadow, and brand style with a single click. Whether you're pointing it at a live URL or feeding it a pile of messy assets, it automatically generates a full, documented UI kit that's ready to hand off as your team's single source of truth. No more chasing original source files; just drop, extract, and start building.

I didn't just want this to scrape data; I wanted it to actually think like a designer.

I built the brain of UIKit by training the AI using Gemini and Claude to deeply identify a brand's DNA. It doesn't just show you colors; it exports full design tokens in JSON format so your developers can implement them instantly. If your source material is missing a button or a specific component, the AI understands the existing logic and generates the missing pieces on the fly. Plus, it handles the vocal language — extracting brand do's and don'ts to deliver high-quality copy for headers and CTAs that actually sounds like the business.

Design and marketing finally speak the same language.

I built an AI sandbox that bridges the gap, using your extracted brand DNA to act as a personality guardian. It's a content powerhouse that lets you generate on-brand copy for any surface.

I started by sketching a futuristic design in Cloud Design and then used roadmap.sh to learn the technical side. The real struggle was in Google AI Studio, where I spent a day training Gemini and Claude to look at screenshots and actually understand a brand's DNA — everything from colors to the way they write. Then used Cloud Code to pull it all together. It was a challenging learning curve, but it turned a messy process into something simple enough for anyone to use.

MetaPrompt: From Messy Ideas to Aesthetic Realness.

Let's be real — prompt engineering is a whole chore. Most of us just want a fire visual without acting like a coding wizard. That's why I built MetaPrompt.

Not your average prompt generator — it's a logic engine.

MetaPrompt isn't your average prompt generator — it's a logic engine. It uses Deconstruction and a Brief Graph to map out your ideas so they actually make sense. Think of it as Proactive Co-creation: it stops you from getting generic, mid results by helping you engineer the prompt like a pro. You're not just typing; you're building with a co-pilot for generation that actually hits.

Once you're in the dashboard, no blank canvas anxiety here. You've got a Pre-made Prompt ready to go — use it as a base or start fresh. The chatbot lets you toggle between generating an Image or a Story, and you can dial in your Ratio and Resolution right there to get the technicals out of the way before the deep dive.

Building this wasn't just about UI.

It was a total technical logic puzzle. I spent days trying to "set the brain" of the LLM to actually deconstruct human thought rather than just spit out generic answers.

My process wasn't a straight line — it went through a constant, circular dance between multiple AI brains. I used Google AI Studio as the primary lab for the "brain" — testing functions and seeing if the MVP could actually handle the logic I was throwing at it. For the visual side, Claude was my go-to for designing components and assets, while Gemini and ChatGPT acted as my constant critics and brainstorming partners. I'd jump between these tabs in a loop: engineering the logic in one, testing the design in another, and then going back to refine and iterate. It was messy, but that's how I managed to turn technical chaos into something that actually works.

Clarifications keep you in the loop, then the prompt updates in real time.

Three steps, one loop: write, clarify, confirm — and the prompt updates automatically after each.

The logic map behind every prompt.

This is the logic map. It shows how every individual, object, and environment detail is linked. Some nodes are solid, some need more energy. You can click any card, edit the info — like atmosphere or settings — directly, and the whole prompt adapts in real time.

Drop a video or GIF

Brief Graph — clicking nodes, live prompt update

Your aesthetic toolkit.

Then we've got the Image Attributions tab. Think of this as your aesthetic toolkit. Use the dropdowns to dial in the location, style, or time of day. It's for those hyper-specific details that make the AI's job easier and your result way more high-end.

Drop image, video, or GIF

Image Attributions — location, style, time-of-day dropdowns

No more mid results — just 100% custom, high-quality visuals.

Once the graph is looking right and the vibes are confirmed, hit Generate. The AI builds out the full imagery and story thread. Done.

A few of the results, straight from the graph.

Drop image

Confessions of a 'Vibecoder': Why I Built AIOS.

I've spent years mastering the 'vibe coding.' As a designer, my world was UI, front-end aesthetics, and creating experiences that looked perfect. But every time I tried to turn those visuals into a living, breathing web app, I hit the same wall. I could paint the car, but I had no idea how to build the engine.

I realized I was caught in a massive skill gap. I had the vision, but backend logic, database architecture, and full-stack deployment felt like a black box.

I found myself drowning in a sea of resources. There are endless AI tools, roadmaps, and tutorials, but they all felt incredibly generic. They didn't know my specific stack, they didn't care about my current skill level, and they certainly didn't provide a real-time GPS for the specific problems I was facing. I didn't need more 'what to build' prompts; I needed to understand how engineers actually think.

Unlike generic AI tools that start by assuming your needs or drowning you in code snippets, AIBOS begins by understanding who you are and how you think.

It doesn't just ask what you want to build; it audits your skills, your stack, and your mental model to create a personalized bridge between your design vision and engineering reality.

By setting my skills and stack upfront, I give the AI context, not just instructions. This stops the drift into generic code and ensures the path is built for my level, focusing on progress instead of fixing misalignments.

Before diving into the code, AIBOS forces me to stop and orient.

It replaces the anxiety of the blank prompt with a guided 'Step Zero' that establishes the specific level of complexity I'm aiming for.

Building the project's 'brain.'

Instead of jumping straight into messy code, AIBOS focuses on building the project's memory first. It transforms the workflow from a linear path into a feedback loop where you're not just 'vibe coding,' but actually thinking like an engineer. By establishing a structured truth through artifacts, you ensure that every response is accurate and every asset is aligned from the start.

Drop image, video, or GIF

recording / prototype embed — the project brain in action

Mochi: The Tamagotchi That Does Your Work.

Emotional leverage meets autonomous utility.

Drop video, GIF, or image

demo — Mochi prototype walkthrough

Digital pets fade. AI agents hide. Mochi does neither.

Instead of one chat prompt, Mochi switches modes.

Presence, personality, and clear boundaries — that's what makes people keep it around.