RimaruBook a call

Want to tell me what challenges you're facing?

Here's what I can do for you, and how we'd start.

Make the work easier to run.

If a process depends on someone remembering every step, I'd start there. My own AI work kept forgetting what the last chat knew, so I thought: what if the repeated steps ran on their own, and a person or an AI only stepped in where something had to be decided? So that's how I map the work now, and then I build the tools and instructions around it.

What that can include

  • Shared project context
  • Review queues
  • Reports
  • Handovers
  • Automations with clear limits

You get work that no longer depends on one person's memory.

How it runs

  • Sort every step.Is it repeated, or does it need a decision? Repeated steps become scheduled jobs that need no AI at all. Decisions go to an AI with written limits, or stay with a person.
  • Show the numbers first.Anything that changes live data gets a dry run, so you see what will change before it does, and a read-back afterwards to prove it did.
  • Write down who may do what.Every AI chat starts from the same context and the same rules, with a clear line between what it can do alone and what needs your OK.

Proof

I run my own work this way. Maru gives each new AI chat the same context, rules and permissions before it touches anything, and I use it every day.

See Maru
A Vizi board in three lanes: a new AI chat reads the map, Maru checks precedence and permissions, the owner approves when asked first, and the chat does the work and logs what changed

Maru as a board in Vizi: a new AI chat reads the map, checks which rule wins, then acts alone or asks the owner first.

Work through a product problem.

I like getting a usable version in front of people early. A request usually sounds like a solution, so I ask what problem it was meant to fix, and then what a person would actually do on the first screen. We work through the journey together, test the parts that are hard, and turn each decision into a prototype or a build handoff.

What that can include

  • Discovery
  • Page and workflow design
  • Clickable prototypes
  • Requirements and acceptance checks

You get a prototype people can click, with every decision already made.

How it runs

  • Start from the problem, not the first idea.Each problem gets written down and checked against what already exists before anything is designed.
  • Draw it with real content.Every screen uses real example text, never placeholders, so you can judge it the way you'd judge the finished product.
  • Hand it over in writing.The engineer gets an interactive design, the decisions, and what done looks like, in the tracker, so nobody waits for a meeting.

Proof

I designed Flomaru's seller dashboard, admin dashboard and seller onboarding this way: every screen drawn with real example content, then handed to the engineer in the tracker as an interactive design.

See Flomaru
Flomaru seller dashboard: today's deliveries at a glance

A Flomaru seller dashboard screen as designed, with real example orders: the kind of screen an engineer builds from.

Find out what the business actually needs.

Sometimes the answer is a build. Sometimes it isn't. When the data told me two different stories about the same customer, I stopped asking what to build and started asking who is actually buying. I'd rather find that out first, keep what the evidence shows apart from what everyone assumes, and tell you to wait than watch you build the wrong thing.

What that can include

  • Customer discovery
  • Market and competitor research
  • Operating plans
  • Financial models built with AI and reviewed by me

You get a written answer you can act on, with every claim labelled.

How it runs

  • Label every claim.Measured, surveyed or assumed. A gap stays a gap, and I don't fill it with a guess.
  • Read each source for what it can tell us.Sign-ups, ad reach and paid orders can describe three different people, so I don't pool them into one.
  • Say what would prove me wrong.You get what we know, what we don't, what I'd do next, and the result that would change my mind.

Proof

At Flomaru, sign-ups and cheap ad reach were mostly men in the Gulf. The people who actually bought were mostly women in Canada. Pooled together, the data described a customer who didn't exist.

See how I worked out who Flomaru's customer is

Connect the product to its customers and partners.

I work through how people find the offer, understand it and get started. At Flomaru I asked what a seller sees the moment they apply, and what happens when one document is missing. That became a seller journey, and it can just as well be a partner program, a campaign experiment or a better way to follow up.

What that can include

  • Onboarding
  • Localized content
  • Partner materials
  • Measurement that tells us what happened without making up the result

You get a path people can follow, and numbers you can trust.

How it runs

  • Walk it as the person living it.In Flomaru's seller onboarding that meant finding the dead ends: a rejection with no way to ask for one missing document, and emails in English for a seller who applied in Spanish.
  • Speak each person's language.The onboarding runs in 13 languages, checked by an AI reviewer panel.
  • Set the rules before the experiment.One channel, one weak spot, a baseline to beat and a stop rule. Money numbers come from the database, not the ad pixel.

Proof

I designed Flomaru's seller onboarding, including the screen where the team reviews each application, and wrote the rules its growth experiments run by.

See Flomaru seller onboarding
Flomaru's admin screen for one seller application: its steps, its answers and the buttons to approve, ask for changes or reject

Flomaru seller onboarding, as designed: "Ask for changes" requests a missing document, so a rejection isn't a dead end.

Teach your team to build with AI.

Most teams get one good answer from AI and one bad one, then stop trusting it. I had the same thing, and I don't think the fix is a better prompt. It's a way of working: small steps, a clear idea of what done means before the AI starts, and something other than the AI checking the result. Maybe that is the part worth teaching, on one real job of yours rather than with slides.

What that can include

  • Planning work in steps an AI can follow
  • Saying what "done" means before the work starts
  • What a program should check and what a person should judge
  • Starter skills and instructions written for your team's work
  • A guided setup for people who are new to it

Proof

I already package my own skills as single files another person's AI can use, and I wrote a guided setup that walks a new user through six steps.

See how I package a system for someone else

Audit what you already have.

Risk in something that already runs is usually a feeling, not a list. I wanted my own company's code to stop being a worry, so every problem got written down on its own, ranked, and handed to the engineers in order. When two numbers disagree, I want to know which one is wrong. Maybe that is all an audit needs to be: a list people can start on.

What that can include

  • Code and infrastructure
  • Tracking and data: which number is true
  • Designs checked against the live code
  • How content is translated and reviewed
  • A ranked repair plan to work from

You get a ranked list your engineers can work through from the top.

How it runs

  • Inspect.AI reads the code and the infrastructure, or the data, and looks for problems. I direct it and decide the order.
  • Write each finding down on its own.So it can be tracked, and argued about, one at a time.
  • Rank, then check the repair.Findings are grouped into stages by seriousness. Engineers fix from the top, and when a fix lands I check that it did what it should.

Proof

I ran this on my own company's code and infrastructure, and it became a ranked repair plan for our engineers.

See how I audit a codebase with AI
A Vizi board: AI inspects the code, findings are written down and repairs ranked, engineers repair, and a check sends unverified outcomes back until the record is updated

Auditing a codebase with AI, as a board in Vizi.

Take a design all the way to a build-ready handover.

A good design can still fail at the handover, when the engineer builds something close and then has to ask what you meant. So I make the handover answer those questions before they're asked. That means every screen drawn, every decision written down, and every button given somewhere to go.

What that can include

  • A brief, rules and decisions kept in files, not in a chat
  • Every page drawn and checked before you see it
  • A frozen design you approve
  • A handover in the tracker, one issue per surface
  • Checks that the build matches the design, screen by screen

You get a design an engineer can build from without a meeting.

How it runs

  • Write the rules down first.The brief, the rules and the decisions live in files before any page is drawn, so the work doesn't depend on what a chat remembers.
  • Check every page twice.A plain program checks the mechanical things. A fresh pair of eyes checks the rest. Both happen before you see it.
  • Hand over in the tracker, not an email.One issue per surface, with every design link, the decisions made and what done looks like. In my own test, readers given only those issues knew what to do 7 times out of 7. Given only the first email, 2 times out of 7.

Proof

I've run this process three times: on Vizi, on Flomaru's dashboards and on seller onboarding. On Vizi it ended in a full redesign handed to my technical co-founder.

See how I redesign a product
A Vizi board in two stages: a brief and an audit, then design rounds the owner approves, then a frozen board, a build and test loop, and a handover through issues

Redesigning a product, as a board in Vizi.

How we'd start.

A first call.

You tell me where the work gets stuck. I ask how it really runs today, not how it's meant to, and who does each step.

What I need from you.

Someone who does the work every day, a few real examples of it going wrong, and a look at whatever exists now: a document, a screen, a spreadsheet, the code.

What you get back first.

A short written note: the problem as I understand it, what I'd check first, and whether I think this needs a build at all. If I've misread it, you'll see that before anything is built.

Want to tell me what challenges you're facing?