AI SYSTEMS

I build the system,not just the prompt.

I work as one person, but there's an operating model behind me. Instead of chatting my way to an answer and hoping it holds up, I built the scaffolding that makes the answer good by default. A library of skills that routes itself to the task, two knowledge sources that keep me grounded, and a set of automations that quietly run on their own. The rest of this page walks through how it fits together.
>
0+
Skills, routed
0
intent buckets
0+
documents I can search
0
knowledge sources
THE INFRASTRUCTURE

The system, powered on

This isn't a shelf of gadgets, it's a live command center. Planning, orchestration, execution, and verification all run on top of a grounded memory system, a self-hosted n8n server, and Zapier when that's the faster route. Turn the dial or tap any subsystem to bring it online.

POWERING UP
Drag the dial or press a node
OPERATING MODEL
Model layers
Plan / Manage / Execute / Verify

Four roles instead of one chat window. Planning sets the strategy, orchestration manages the work and holds the gate, execution layers run in parallel, and an adversarial plus live-runtime layer verifies on the real surface. Memory keeps all of them grounded.

4roles in the operating model
THE OPERATING MODEL

A studio of one, run like a team

I treat models as roles rather than one big chat window. Planning owns the strategy, a manager sequences the work and holds the quality gate, the workers build under those briefs, and then I verify by driving the real thing myself. That structure is what lets one person carry the load of a small team without letting the quality slip.

01Plan
Planning layer

This is where the strategy and the pipeline architecture get worked out. It decides what gets built and why, before a single line is written.

02Manage
Orchestration layer

Briefs get written down to specific values here, the work gets sequenced, and nothing moves past this layer until it clears the gate.

03Execute
Execution layers

The build itself runs fast and in parallel, strictly under the brief, because every hard question was already settled upstream.

04Verify
Verification layer

Reviewers fan out, I argue with every finding they hand back, fixes get ranked by what actually matters, and then I go and drive the real surface.

A passing log isn't working software. I don't call anything done until I've driven the real surface myself.

CAPABILITIES

What I actually do with it

The tooling only matters because it's pointed at real work. Across design, code, research, product, content, and automation, AI does the heavy lifting while the judgment stays mine.

C01
Design and frontend

I build production interfaces and the motion inside them, then say what I honestly think before any of it ships.

C02
Build and ship

Real code on a stack I know cold, plus the hosting, the deploys, and the check on the live product afterwards.

C03
Research and evaluation

When an answer has to be right, I go and find it, then build the checks that keep the output honest.

C04
Automation

Workflows that move data between the tools a business already pays for, quietly, without me in the loop.

C05
Knowledge systems

Two independent sources feed real context into every task, so nothing has to start from a generic guess.

C06
Media and audio

Local pipelines clean up audio, transcribe it and label who spoke, and I ship them as tools I can reuse.

I'm not a data scientist and I don't train models. What I am good at is building the system around them, so what comes back is reliable, on-brand and fast.

SKILL LIBRARY

A routed library, not a folder of prompts

Anyone can collect prompts. The hard part is reaching for the right one every single time without stopping to think, so I built a router for it. Every task gets parsed, sorted into one of eighteen intent buckets, and matched to the skill that fits. If two skills overlap, a rule settles it, and when it's genuinely a toss-up, a small ranking script scores the candidates and hands back the top three, so the pick stays repeatable instead of a coin flip.

PARSE
SCANNING 18 BUCKETS
90+ skills | 18 intent buckets | 8 domains

The whole thing regenerates itself. I add a skill, rebuild the index, and the router already knows about it, so it gets a little sharper every time I use it.

MEMORY

A brain that stays grounded

A wrong answer sounds just as confident as a right one, so nothing here gets remembered on a hunch. Raw inputs are written down once and never edited, and a distilled wiki sits on top of them, citing its source for every claim. A local engine indexes the whole lot, and my skills read that context before they generate a single word. Two independent sources feed the same five layers: my own craft, and whatever client work happens to be live.

APPEND ONLY | 3,300+ DOCUMENTS | FIVE LAYERS
HOVER A RECORD | HOLD TO SLOW
01Raw inputs
02Distilled wiki
03Index and history
04Local search
05Skills read it
Two independent sources sharing one structure, and neither one writes into the other.

One night I let the system run an autonomous pass over itself: audit, build, verify, fix, across four waves and well over a hundred agents, and every page came out carrying an honest note on whatever was still unverified. I would rather it be grounded than impressive.

AUTOMATIONS

Work that runs without me

The best automation is the one you forget is even there. I host my own n8n server, and separately I will reach for Zapier when that is the faster path. Forms fill the right lists, payments land where they belong, and follow-ups go out on schedule without anyone copying data between tabs.

LEAD ROUTING
Every form lands in the right list

A single webhook catches every form on the site, stamps each submission with the page it came from, and files it into the right audience list, so I never sort leads by hand.

/pricing | filed to Waitlist
PAYMENTS TO CRM
Payments file themselves

Payment activity writes itself into the marketing platform, split so paid, failed, refunded, and disputed each land in their own list with the order details attached. Nobody has to open the payment dashboard.

Each outcome to its own list | counts start at zero on load

Past these two, I've built messaging automations on the WhatsApp Cloud API, stood up a self-hosted server on my own infrastructure at my domain, and added safeguards that catch gaps in a sales pipeline before they cost a deal.

0
live cloud automations running
0
automation server I host myself
Zero
tabs copied by hand
SETUPS

The environment I built to move fast

Speed comes from the setup, not from typing faster. My environment connects the models directly to the tools I use, so the same assistant that plans a project can edit the code, drive the browser, query the knowledge base, and push the deploy. Every recurring correction becomes a durable rule, so the system stops making the same mistake twice.

U01
Connected toolingThe models plug straight into the tools I design, code, browse, and host with, so one flow carries a project from plan to ship.
U02
A skill for every jobMore than ninety routed skills sit ready, so the right method loads on its own for whatever's in front of me.
U03
Grounded by defaultMy knowledge sources and local search feed real context in before a task even begins.
U04
Corrections stickEvery correction gets promoted into a permanent rule or a taste note, so the quality keeps building on itself.
U05
Verified on the real thingNothing ships on a passing log alone. I drive the live surface and check it with my own eyes.
WHERE I PUSH IT

Ideas I chase on my own time

I don't wait for a brief to experiment. When a new capability shows up, I build something real with it just to find where it breaks. Some of it ships, some teaches me a lesson and gets shelved, and all of it feeds back into how I work.

01SHIPPED
Opening up what is reusable

When a personal tool is good enough to help other people, I clean it up and ship it publicly. My audio pipeline for cleanup, transcription, and speaker labeling went out as a real open-source project with its own tool integrations and a full test suite.

02LIVE
The toolkit is a product

I keep a live gap analysis of the wider tooling market against what I already run, so I can see what is genuinely worth adding instead of chasing every release.

03ONGOING
Trying things before I adopt them

New frameworks get a real trial instead of a hot take. I install them, push them hard, and write down an honest verdict on whether they have earned a place next to what I already run.

04ACTIVE FOCUS
Motion is where the craft is now

The interfaces I care about most are the ones that move well, so motion is where most of my craft goes these days. I check every animation by driving it rather than assuming it works.

WORK WITH ME

What I do for clients

The same operating model runs on client work, across three lanes. I bring AI in where it actually earns its place, design that can carry enterprise weight, and development that ships the result as real, live code, all of it handled by one person from start to finish.

SVC-01
AI
AI systems and products
The systems on this page, built for you.
AI product designAgent BuilderChatbots + integrationsWorkflow automation

I built a working AI chatbot for a real-estate platform and run live automations on my own n8n server and Zapier.

SVC-02
DESIGN
Product design, 0-to-1
Enterprise-grade UX, without the wait.
0-to-1 SaaS, fintech, enterpriseUX and UI designDesign systemsResearch + CX / service design

Enterprise UX for teams at Microsoft, SHL, and Arkk, plus 0-to-1 products taken from a blank canvas to launch.

SVC-03
DEVELOPMENT
Build and front-end
Design that ends up as live code.
Framer + Webflow buildsFront-end engineeringLive interactive componentsMotion + micro-interaction work

This site, its live coded components, and client sites are built and deployed by me under my studio, DROPCAP.

If you can describe the system, I can build it.

Maybe it's an AI-assisted product, maybe a knowledge base that stays honest, maybe automation that quietly handles the boring parts of a business. Whatever the shape of it, I design the whole system around that goal and ship it.