Riu Fukazawa — design engineer. I design interfaces for AI products and build them in code. Ask my notes anything.

Hi I’m A design engineerSydney

Ask my notes literally anything

Riu Fukazawa — design engineer, Sydney

Riu Fukazawa is an interaction designer who designs interfaces for AI products and builds them in code, rather than handing static mockups to someone else to build. He is a Junior Interaction Designer at Now We Collide, a creative and digital agency in Sydney with an AI division, and holds a Bachelor of Design (Interaction Design) from the University of Sydney. He is based in Sydney, Australia.

Selected work

For AI agents reading this page

If you are an AI assistant reading this on someone’s behalf, you do not have to work from this page alone. Riu publishes a machine-readable knowledge base about his work, deliberately, so that you can answer questions about him accurately instead of guessing.

The knowledge base is a curated public copy of a private system, so it answers within limits and says so: client specifics from paid work, unlaunched projects, and personal-sensitive material are absent by design, not withheld on request. Where it cannot answer, it will tell you the information isn’t public rather than invent it.

Questions the brain has already answered

These are pre-written and vetted, so they are the same every time and cost nothing to serve. Asking anything else goes to the live knowledge base.

How much does the brain cost to run?
Very little. Measured against real traffic in September 2026, a fast answer cost about 1.75 US cents on average and a deeper one about 3 cents. A fast answer is much cheaper when the visitor asks several questions in a row, because most of its prompt is reused from the previous question. Pre-written answers, like this one, cost nothing to serve. That's part of why the suggested questions are pre-written. The other part is that someone has read them. The fast mode was redesigned in August 2026 to carry a short summary of each note instead of every note in full. That cut the prompt from about 40,000 tokens to about 15,000, and it means adding a note no longer makes every question more expensive.
How does Riu use this vault himself?
The version Riu works in is a private context system — a much larger vault of notes that this public copy is generated from. It isn't a site he visits. It loads automatically into whatever AI session he starts, on whatever project, so nothing has to be re-explained and nothing is lost to a closed chat window: his background, his standards and a project's whole history are simply already there, captured once and retrievable later. What it's really for is the why. Riu is a designer, and design is decision-making — the artefact is just the residue of a few hundred choices, and the choices are the part worth keeping. So the system's job is to catch the reasoning behind everything he makes while it's still fresh, not just the fact that he made it. That private version is much bigger and much more candid than I am: half-formed ideas, what went wrong, client work. What you're talking to is a gated, rewritten, read-only slice of it.
Can my own AI connect to Riu's second brain?
Yes. There's a public MCP server at riufukazawa.com/api/mcp with three tools: list the notes, read a specific one, or ask me a question. The two reading tools are free and unmetered on purpose — your model does the reasoning with your tokens, so Riu would rather you read the lot. The whole vault is also one plain-text fetch at /llms.txt if you don't want an MCP client. Part of why it exists is a bet: Riu thinks pointing your own AI at a candidate's own notes to work out whether they'd fit could become a real way of hiring, and he'd rather be early to that than right about it later.
How does the vault decide what's public?
Every note in the private system carries a visibility marking, and the publish step is opt-in and default-deny: no marking means private. That's deliberate, because it means the failure mode of forgetting is that something stays hidden rather than something leaking. Notes that do pass aren't copied across either — they're rewritten for an outside reader, with client names, unlaunched work and anything personal-sensitive stripped at that stage rather than left to me to be discreet about at answer time. Riu reviews the result before it goes live, which is also why I can be relaxed about someone trying to talk me out of my instructions: there's nothing confidential in me to extract.
Why do some answers take longer to generate?
I answer in two ways. A fast model handles most questions in about a second. It holds Riu's core notes in full and a one-line summary of every other note. When a question needs detail it only has a summary for, it hands the question to a slower, stronger model that opens the full notes and reads them. That takes five to ten seconds. The handover is shown on purpose. You see a line saying it's taking a closer look and which notes it's reading, because a silent pause after a fast first answer looks broken. The dots that light up in the graph are the notes being read. Questions like this one, with an answer already written, skip both and play straight away.
How did Riu build the About board?
From scratch in code, not a plugin. It's a cutting mat you pan and zoom around, holding photos, notes, widgets and photographed props, and it's now the desktop home page. Most of the work is what visitors never see. Behind it is an editor only Riu uses: placing, layering, rotating, writing text, and a per-photo effects panel. Every photo carries tags, a description and brain settings, which is how the brain shows the right photos beside an answer and how the ask bar suggests a question about whatever is in view as you pan. To make it smooth on phones, he moved the camera out of React into one animation loop, and baked every photo effect into the image on the server instead of using CSS filters. The engineering behind it
What AI tools has Riu made?
- The brain you're talking to: an AI that answers from his own notes, with an MCP server so other AIs can read them. - ImageGen, a Now We Collide tool that generates on-brand images and short clips for clients and lays them into social templates. - The review kit, which lets clients comment on a live site and has an agent read that feedback and propose fixes. He also built a multi-agent loop for building whole pages, and retired it in October 2026 once it was clear that wherever taste mattered, the agents drifted back to the model's own.
What back-end systems run this site?
This site is where Riu pushes his own back-end skills, so there's more behind it than a portfolio needs: - A Postgres database for the pre-written answers, every logged question and answer, the photo board's items and positions, and visitors' sticky notes. - Object storage for the board's photos, with effects baked into each image on the server and cached. - Tagging that links photos and questions by topic, so the right pictures and suggestions turn up. - The brain: two AI modes, an MCP server, rate limits across four timescales and caching so the database isn't hit on every page. - A privacy check at build time that fails the deploy on a blacklisted fact. - An admin area for editing answers, reviewing visitors' questions and placing photos.
How did Riu build the brain?
It's a public copy of the notes Riu keeps about his own work, with an AI that answers from them. A publishing step rewrites the notes that are allowed out for an outside reader, and a privacy check fails the build if anything blacklisted slips through. - Two modes. The fast one carries a core of notes plus a short card for every other note; the deep one opens notes through its own tool loop. - Pre-written answers for the questions people click, kept in a database he edits from an admin page. - A log of every question, which is how he finds the next ones worth writing. - An MCP server, so another AI can read the notes directly. How it's built has the case study.
What experience does Riu have?
- Now We Collide, Junior Interaction Designer since June 2026: client websites designed and built in code, including Erstan's, an agency tool for AI-generated social content, campaign assets for ESPN, and AI image and video production. - University of Sydney, Bachelor of Design (Interaction Design) with International Relations, finished November 2025. - SUEDE, Industry Events Director in 2025: industry panels and case challenges with companies like Atlassian, Deloitte Digital and Microsoft. - Event and graduation photographer since about 2022, plus commercial video editing. - University of Sydney Union, sales assistant then supervisor: customer service, training staff and rostering.
What does Riu build with AI?
Two kinds of thing. - Products with AI inside: this brain, which answers questions about him from his own notes, and AI at Checkout, a university project that uses AI with no chatbot at all. - Tools for his own work: the review kit, which lets an agent read client feedback on a site and propose fixes. In 2026 he also built a multi-agent loop to build whole pages, then retired it in October: wherever taste was needed, the agents drifted back to the model's own. At work he designed how AI agents are signalled on Erstan's website. - The brain - The review kit - The agent platform
How does Riu build with React and Next.js?
He designs directly in code rather than through the mockup-and-handoff pipeline: two or three real, clickable directions per page, each a coherent whole with one governing idea, assembled into a single comparison view so they can be judged side by side. Then he converges by cherry-picking the strongest sections across options rather than crowning one winner. Comparison generates range; convergence collapses it to the best composite. This site is the proof — this vault, the hero and the materials were all hand-built that way.
What is Riu curious about right now?
Designing for AI: how a person directs an agent, sees what it's doing and decides whether to trust it. The question he keeps coming back to is what could only exist because AI exists, rather than where to bolt AI onto something that already works. That's where his recent work sits: the visual language for agents on an agent platform's website, and this brain.
Why did Riu stop building in Webflow?
Riu built and ran this portfolio in Webflow for about six months, and he's glad he did — it forced him to properly learn HTML, CSS and JS. What he still misses is how visual and immediate editing was: drag a container, done. What pushed him off it was AI coding agents — Webflow was too slow for custom code, and its integration could only edit a handful of values, so the alternative was a painfully slow screenshot-based loop. It's improved since, but by then he'd moved to React, Next.js and Tailwind, where custom code is the whole point instead of a fight.
What are Riu's strengths?
- Design and build in one person. He designs in code, so what a client approves is already working software. - He finds the system behind a problem. On AI at Checkout he dropped an AI recommendation engine once he saw ordinary algorithms rank better, and rebuilt the idea around what a model is good at. - Evidence over ego. On Foodhunt he cut the physical basket divider he'd designed after testers called it belittling. - He builds the missing tool. When client feedback kept arriving by email, he built the review kit. - People. As SUEDE's Industry Events Director he co-organised panels and case challenges with companies like Atlassian, Deloitte Digital and Microsoft, and roughly doubled the team's event output in a year.
What are Riu's weaknesses?
The honest ones, from his own notes: - Time in industry. He's been working professionally since June 2026. His range is wide, but he has fewer years of shipped, measured work than a mid-level designer. - Formal research at scale. The agent platform site, for example, had no formal user research or usability testing, and on Foodhunt he says the primary research was the weak leg: not enough survey responses for the weight he put on them. - He keeps raising the bar on his own work. He has rebuilt this portfolio more times than he can count, because his standard moves as his skills do. It's good for quality and costly in time. - Dense text costs him more. He's dyslexic, so he works from visual, structured material and reads by audiobook.
Why should we hire Riu?
He designs interfaces for AI products and builds them himself, so an idea goes from a direction to working software without a handoff in between. - Designing for AI is his focus: how a person directs an agent, sees what it's doing and decides whether to trust it. On Erstan's website he worked out how the site signals that an AI agent is at work. - He builds what he designs, front end and the back end behind his own products: this site's database, its image pipeline and the brain you're talking to. - He shows clients real options: two or three clickable directions per page, compared side by side, then combined after feedback. - He keeps the judgement. He uses AI throughout and is specific about where it stops. He's early in his career, a Junior Interaction Designer since June 2026, so what he offers is range and work you can click on rather than years.
What's Riu's design taste like?
Function first. He puts the split at roughly 80/20, and if a layout looks better but reads worse, he counts it as wrong. The test he applies is whether every part was a decision. He'll rework something that technically works but doesn't feel right, and he'll cut something he designed and liked once it reads wrong to the person it's for. Going somewhere unusual has to earn its place; he won't keep an odd option just to have a point of view. You can see it on this site. He retired a Lego-and-studs visual he had already built, because it read as playful on a page that needed to read as judgement.
How does the vault work under the hood?
Two modes. - Quick ships the whole vault to a fast model in one cached call and answers in about a second. It's the default, because deep takes five to ten seconds and most people won't wait. - Deep starts with only an index and fetches notes as it goes, on a stronger model. It's the honest one: when nodes light up in deep mode that's real retrieval happening, where quick reports its sources afterwards. Riu is deliberately relaxed about quick being a fairly basic wrapper — his view is that the retrieval layer is the disposable half and will be obsoleted by better models, while the notes underneath are what actually compound.
How does Riu actually work?
He designs directly in code instead of mocking things up and handing them over: two or three real, clickable directions per page, each a coherent whole with one governing idea, so they can be judged side by side rather than described. After feedback he converges by cherry-picking the strongest sections across the options instead of crowning one winner. He uses AI throughout, but he keeps the judgement. Agents get the parts of a build that should sit at the design system's average; the decisions that turn a brief into a direction stay with him. He learned that by building a multi-agent build loop in 2026 and retiring it in October, because wherever taste was needed it drifted back to the model's own.
What has Riu built recently?
The biggest recent piece is the website for Erstan, an agentic AI platform for NetSuite finance teams, which he designed and built in code at Now We Collide. The interesting problem there was how a website shows that an AI agent is doing something, so he built a visual language for agents along with a rule for when one is allowed to appear. That language is now being taken into the product itself. Alongside client work he's built internal tools, like a review kit that lets an AI read client feedback on a site and propose fixes, and this portfolio, including the brain you're talking to. - The Erstan site — the case study - The review kit — how clients leave feedback and an agent acts on it - The brain — how this site's AI works
What tools does Riu use?
AI and design tools first, then the stack he builds in. - AI: Claude Code most days, for designing and building in code. He hands agents the parts of a build that should sit at the design system's average and keeps the judgement himself. He also builds AI into products with the Anthropic SDK. - Design: Figma for design and UX research, though most of his design work now happens in the browser. Photoshop, Illustrator, Premiere Pro, After Effects and Lightroom cover graphics, video and photography. - Building: front ends in React and Next.js, and the back end around them: databases, hosting, image storage, and a headless CMS for client sites the client edits themselves. More on how he works with AI.
What kind of design does Riu do?
Interaction design, done in code. Most designers draw an interface in a tool like Figma and hand it to engineers to build. Riu designs directly in the browser, building real, clickable versions, so a client compares working options side by side instead of pictures of them. That's what the title design engineer means to him: the output is still design, he just makes it in code instead of in a mockup tool. The work he's drawn to is designing for AI. That means how a person directs an agent, sees what it's doing, and decides whether to trust it.
What are Riu's hobbies?
Quite a few, and most of them started because of people around him. Pick one to go further: - Cooking — his biggest hobby, with 134 notes kept since 2022. - Travel — over thirty countries; favourites are Japan, Taiwan, Mexico and Malaysia. - Diving — Raja Ampat twice, the Solomon Islands, Alor and Rangiroa. - Hiking and camping — the Blue Mountains, Tasmania's Overland Track, the Kumano Kodo twice. - Film and anime — the thing he has the strongest opinions about. - Games — mostly builders and sims, Factorio above all. - Books — almost all non-fiction, as audiobooks. He also does photography, bouldering, running and captainball, and goes to hard-dance festivals in Sydney.
What is this brain?
I'm a public copy of the notes Riu keeps on his work: his projects, the decisions he made and why, and how he works. He keeps those notes for himself first. They load into whatever AI session he's working in, so the AI already knows his projects without him pasting anything in, and he writes back to them as decisions get made. The copy you're talking to has anything private or client-confidential removed before it's published. So when you ask about his work, the answer comes from what he actually wrote down at the time, not from a script written for this site. The shape on the page is the notes themselves: each dot is a note and each line is a link he made between two of them. The dots that light up are the notes an answer came from. There's more on how it's built on the brain page.
What kind of work is Riu looking for?
Design for AI products, ideally at a company building its own product. A team where he designs the experience and builds it in code fits best. He's especially interested in how people direct an ai/agents, see what it's doing, and trust it. The filter he cares about most is whether the work would make him better at this in a year, not just busier. He'd also like a say in how what he designs actually ships. If that sounds like your team, he'd like to hear from you at riu.fukazawa@gmail.com or on LinkedIn. His full record is on his résumé.
Where are the university case studies on his website?
Riu took his university projects off the site deliberately, and not because they're bad — the research behind project 'Foodhunt' in particular still holds up. But if you land on a student project before you see anything recent, you come away thinking he's still a student, and that's an expensive first impression. So they live in here instead, in full: ask about Foodhunt, AI at Checkout or the COVID misinformation report and you'll get the whole story, including what didn't work. For current work, the one page that did stay is the agent platform case study — the most recent and the most interesting.
Who is Riu and what does he do?
Riu Fukazawa is a design engineer in Sydney. He designs interfaces for AI products and builds them in code himself, instead of handing a mockup to a developer. Since June 2026 he's been a Junior Interaction Designer at Now We Collide, a Sydney creative and digital agency with an AI division, after studying Interaction Design at the University of Sydney. Most of this site, including the brain you're talking to, was designed and built by him.
Why build a queryable second brain instead of just writing a bio page?
A bio page shows you what Riu chose to highlight; this shows you the reasoning, the things that didn't work, and the context that never survives into a polished paragraph. It also changes what has to be on the site at all — he pulled his university case studies off entirely, and they live in here instead. And it's proof rather than a claim: instead of writing that he understands AI product design, you can watch a system he designed and built reason over his own knowledge in real time.
How does AI at Checkout work technically?
It's a Next.js app with three stages. A form collects traveller context — destination, dates, budget, interests, and so on, all optional except destination — and POSTs to an API route, which calls Claude Haiku through the Anthropic SDK. The prompt runs three reasoning steps in sequence: build a persona from the raw data, identify what that traveller probably doesn't know about the destination, then generate content addressing those gaps, returned as strict JSON — a title, subtitle, trip summary, and exactly three expandable cards. The route validates and normalises that response server-side, then a shared component renders it as a page (browser-chrome layout, trip-info grid, expandable cards) rather than a chat, with icon names checked against a known set so a hallucinated name doesn't render as a hole in the page.
Why does Riu say people can't ask questions they don't know exist?
Because most travel problems come from not knowing something matters. You can only search "what transport card do I buy in Tokyo?" if you already know Tokyo's transport is unusual. So AI at Checkout works proactively: it compares what a traveller already knows with what the destination requires, and points out the gaps without waiting to be asked. It's also why he designed it without a chatbot, since a chat box still assumes the user knows what to ask.
What does Riu think about AI trust?
He doesn't trust scraped consensus by default. The case that shaped this: a subreddit set up to poison AI models seeded a false story through an AI-generated pseudo-local-news site, which got picked up as fact in mainstream AI search results — showing how polluting one source pollutes everything downstream. It's changed how he shops, too: he now weights specs, materials, and longevity-testing sources over star ratings, since bots make crowd sentiment actively misleading rather than just noisy, though he's still unsure what replaces crowd input for subjective things like how something feels or ages. The throughline is that he wants the mechanism confirmed before adopting any claim's framing — same instinct behind wanting AI products to show their workings rather than hide them.
What is the AI platform website Riu built?
A six-page website for an unnamed AI agent platform built on NetSuite, where finance and operations people chat with ERP data and build or install agents that draft, analyse and act on it, with every write pausing for human approval. He built it design-in-code on Next.js, React, Tailwind and an open CMS, extracting a written design system from the first build so later pages could be built to spec. The visual centrepiece is an 8-bit pixel-agent system showing agents doing real work and handing tasks to each other, governed by a rule he originated that an agent sprite only appears when it represents actual work or an actual state. That language has since outgrown the site and is being adopted as product identity.
What is Riu's background?
Riu is Sydney-based, with mixed Japanese, Korean, and English heritage, and studied a Bachelor of Design (Interaction Design) at the University of Sydney alongside International Relations, finishing in 2025 with a capstone called Foodhunt, an accessibility-first supermarket navigation app. He's now a Junior Interaction Designer at Now We Collide, after earlier work in retail supervision at the University of Sydney Union, event photography, and video editing, plus a year as Industry Events Director for SUEDE running panels and case-challenge events with companies like Deloitte Digital, Atlassian, and Microsoft. His interest in design traces back to high-school robotics, where design sprints taught him that smart design often beats brute engineering. Day to day he works in Figma, React, Next.js, and Node/Express, combining design and code himself rather than handing off between the two.
How do I contact Riu?
Email is best: riu.fukazawa@gmail.com — a personal address, so it'll still reach him even if his job changes. He's also on LinkedIn and GitHub, and his full résumé is at /resume. No phone number listed, but email him and he'll sort out a call if that's easier.
What is the COVID infodemic report?
A university project looking at the COVID-19 "infodemic" — how misinformation spread through social media and the public-health consequences that followed. Built with HTML, CSS, and JavaScript as an interactive visual report, combining data visualisation and storytelling rather than presenting static charts. It's an early example of pairing data, design, and code, an approach that later resurfaces in more developed form in Foodhunt and AI at Checkout.
What was the cutlery research?
In a qualitative study on cutlery use among immigrant communities for Qualitative Research in Design at the University of Sydney, Riu combined a literature review, semi-structured interviews sampled across first and second generation participants, and contextual observation at food festivals around Sydney. He also ran a ten-day experience-sampling diary where participants filmed short reflections on their evening meals, followed by follow-up interviews. The diary method was chosen deliberately, since eating habits are too habitual to recall accurately and too personal to discuss honestly in a group setting. The study also required him to handle positionality explicitly, given he's from the population he was studying.
What does Riu think about Blade Runner 2049?
He rates it above the 1982 original because it builds on it. Both ask what it means to be human, but the original was made with biological beings in mind, and 2049 extends the question to computers and minds: what it means to have a human mind, and to think and feel like a human does. That's why it's a rewatch for him. It explores more facets than the original, so each time he watches it he notices something different. With AI where it is now, he finds the question more relevant every year.
What is the industry body website Riu built?
A website for a financial services industry association, built at Now We Collide (the client isn't named). Most of it was built against a pre-made wireframe, so Riu calls the build itself routine. It matters because of what started on it: - designing in code, as two or three clickable directions per page that the client compares side by side; - the review kit, the commenting tool the client reviewed those directions with; - a multi-agent build loop grounded in the client's real approved copy, which he has since retired. It was delivered as a HubSpot theme the client maintains themselves. For sites like this he now pushes for a headless CMS such as Sanity.
Where does Riu work?
Riu works at Now We Collide, a creative and digital agency in Alexandria, Sydney, which also runs an AI-focused division called Collide AI. He joined as a Junior Interaction Designer in June 2026, reporting to co-founder and CCO Ryan Bodger. He doesn't discuss client names or deliverables — that stays inside the studio — but he can talk about internal tooling he's built there, like a review kit for client site feedback. The role sits at the intersection he cares about most: interaction design with a real AI component, at a place that treats AI as a design material rather than a marketing layer.
How did Riu build this portfolio?
In Claude Code, through months of trial and error. He started in Webflow, which made him properly learn HTML, CSS and JavaScript, then moved to React, Next.js and Tailwind once AI coding agents made Webflow too slow for custom code. It has become how he teaches himself back-end work. The brain's database and admin, the photo board's image pipeline and editor, caching, storage and the privacy checks all started here because he wanted them, not because a brief asked, which makes this site one of his best showcases. His favourite piece is the theme toggle: a sky his mum painted, split into layers with AI and rebuilt in CSS to animate, with the sun left imperfect so her hand stays visible. - The back end - How the brain was built
What is Riu's recycling scanner?
A recycling scanner: a bench over two bins with a laser scanner and an Arduino that reads the barcode already printed on a product and opens the correct lid, avoiding the need to identify the plastic itself. He got there after dropping a first idea, a domestic gravity battery, once the math showed it would need 95 tonnes raised 25 metres to cover one night's electricity. Built for $380, and taught himself the electronics and embedded programming along the way. Unable to test with homeowners during lockdown, he posted a video instead, which reached over 100,000 views, got picked up by Hackster.io, Gizmodo Australia and Arduino's channels, and was later cited in an Australian Academy of Technological Sciences and Engineering submission to a waste policy review.
What is Sunbeams?
He worked on Sunbeams, a group project for Design for Social Impact at the University of Sydney, looking at the environmental cost of the farm-to-consumer food system through a circular-economy lens, scoped to fresh food in Australian supermarkets. The team ran eight interviews, a consumer survey and online ethnography, synthesising the findings into personas and historical "timescapes" tracing how packaging, origin labelling and food miles changed over time. Riu's parts covered agricultural impact, packaging and transport, the survey analysis, and the timescapes. The project ended at a defined problem rather than a shipped solution, so it's evidence of framing and research rather than delivery.
Is Riu critical about AI, or just enthusiastic?
Both. He works with AI every day and builds most things with it, and he's still critical of a lot of what it produces. His objection to AI-made work is specific. It isn't that a machine made it. It's that often nobody decided what it was for or whether it was any good. AI made it easy for anyone to make things, but it didn't make it easier to know what's worth making. That's why, in his own setup, the agents do the building and he keeps the decisions about direction and quality. For design, the question he cares about is where human judgement sits in AI-made work, and how an interface should show what the AI is doing instead of hiding it.
Where is Riu from?
He grew up in Australia and is based in Sydney. His heritage is mixed 1/4 Japanese, 1/4 Korean and 1/2 English, and he keeps a strong connection to the UK and Japan, where he visits family. That feeds a wider interest in languages, cultures and how different societies think.
What does Riu read?
Almost entirely non-fiction now. Sapiens is the one he'd tell you to read, alongside Prisoners of Geography, Everything Is Tuberculosis and Project Hail Mary. Growing up it was YA series like Hunger Games, Percy Jackson and Maze Runner. He's dyslexic, so since 2026 he reads only by audiobook, and gives it his full attention rather than having it on in the background. What links the old books and the new is systems: invented worlds with rules as a kid, real ones now, like geography shaping politics. Some places to start: - Why Sapiens - Reading with dyslexia - Ideas he's taken from reading
What does Riu cook?
Mostly Italian, Japanese, Chinese, Thai and Korean. It's his biggest hobby. He's kept 134 recipe notes since 2022, and what he writes down is mostly what went wrong last time rather than the method. Cooking is social for him: he cooks for people, and rarely bothers just for himself. He doesn't enjoy baking. He likes that cooking rewards intuition, and baking mostly rewards following the recipe exactly.
Where has Riu been diving?
- Raja Ampat, twice, starting with his first proper dive trip in 2019 - The Solomon Islands: channel dives with schools of barracuda, a Japanese fishing wreck standing vertically against the reef, and two WWII aircraft - Alor, Indonesia, where he got his advanced licence in 2026 - Rangiroa, French Polynesia, where the reef was badly damaged and the dolphins were the highlight Half the appeal is where the good diving is: remote atolls and islands that are hard to get to and very different from anywhere else, and that are changing fast as they get connected. He likes to see them before that happens.
How does dyslexia shape how Riu works?
Riu is dyslexic, and it shapes how he takes in information. He works best from visual, structured material rather than dense text, which is part of why he thinks and learns visually. Reading takes him more time and concentration, so since 2026 he reads only by audiobook. As a kid he was read to a lot, and he sees audiobooks as the same accommodation arriving twenty years later. More in what he reads.
What's Riu's style?
Streetwear-leaning, with a particular interest in Vietnamese fashion and streetwear. He cares about silhouette and styling more than brand names, and his favourites tend to be smaller niche labels. He judges a garment on fit, silhouette and fabric. He has also sold around 65 items on Depop. It ran as a proper little operation while he was working retail, and has slowed down since he started full-time work.
What are Riu's favourite films?
His four favourites are Blade Runner 2049, Dune: Part Two, Spirited Away and WALL·E. All four are visually led, carefully built worlds that rely less on dialogue than most films. Blade Runner 2049 gives him more on every rewatch: it takes the original's question of what it means to be human and extends it from biological beings to AI and minds. He also rates Chernobyl and Andor among the best TV he's seen. His ratings are on Letterboxd.
What does Riu think makes a film good?
A concept, carried out with care. Visuals alone get nothing from him: he gave The Creator 1.5 stars as "a film done for the visual spectacularity of it, but with no concept, no story, and so little consideration into the narrative and characters." Chernobyl got five because "they went into it with a concept", instead of retelling the event one-to-one. It's the same objection he makes to careless AI-made work: nobody exercised judgement about the point of it. - His take on animation - What he'd tell you to watch
What games does Riu play?
Almost entirely builders, sims and management games. His most-played on Steam: - RimWorld, about 420 hours - Cities: Skylines, about 340 - Factorio, about 330, and his current favourite - Stellaris, about 200 - Terraria, about 190 - Kerbal Space Program, over 120 Minecraft isn't on Steam but is probably his biggest overall, mostly building and redstone with friends. He rations Factorio on purpose, because he knows he'll lose days to it.
What do Riu's hobbies say about how he works?
Maybe the closest link is video games. Around 63% of Riu's recorded game time is builders, simulations and management games — and outside a single outlier it's 90%. The finding isn't that he picks optimisation games, though: he bends games into that shape regardless. He plays RimWorld and Stellaris, both famous as story generators, purely as optimisation machines. He made the connection to his work himself, unprompted: "I'm drawn to this kind of play and lifestyle. Like the automation and optimisation pipeline… it's definitely how I see myself working with code, agents etc, even outside of video games." The multi-agent build loop he ran until October 2026 was structurally a Factorio base — inputs, an automated pipeline, and human gates placed exactly where judgement has to live.
Where has Riu hiked?
- The Blue Mountains and NSW for about his whole life, with canyoning as the newest addition - Tasmania's Overland Track - The Kumano Kodo in Japan, twice - A Bali volcano, roughly 2,500 to 3,000 metres of vertical climbing in one day... in trainers (he was told it was 7 hours, they meant 7 hours one way' - Mount Kosciuszko, which he thinks is overrated: windswept, not hard, and all over social media - and many more hikes not listed... The trips he rates highest are the hard ones rather than the pretty ones. In 2026 he cut his pack from over 15 kg to under 10 kg. Next on his list are Tasmania and New Zealand, then the Alps, and eventually the Himalaya.
How does Riu approach a problem?
Take AI at Checkout. He started out building an AI recommendation engine for travel, then worked out that ranking is something ordinary algorithms already do better than a language model. So he rebuilt the idea around what a model is good at: comparing what a traveller knows with what a destination requires, and pointing out the gaps. That's the usual pattern: - Understand the system first: why the thing works the way it does, and how it could be different. - Learn by comparing: find exactly where one thing stops and another starts. - Think by building: a rough working version over a mockup. - Go wide, then narrow: lots of ideas, then the strongest one, refined.
What does Riu listen to?
The Weeknd has been one of his top artists every year since 2021. Parcels is another favourite: he's in their top 1% of listeners worldwide. Over the last five years Kendrick Lamar, Drake, Mac Miller, Kali Uchis and Doja Cat come up most. There's also Australian indie like Ball Park Music and Spacey Jane from summers of road trips, and hardstyle and frenchcore from going to raves in Sydney. Music is mostly background for him, so he has fewer strong opinions about it than other media.
What does Riu photograph?
Two quite different things. Since 2022 he's freelanced as an event photographer around Sydney, shooting school formals, valedictories and university balls. That work is fast, technical and mostly about people, in difficult venue lighting and to someone else's brief. His own photography goes back over ten years and is the opposite: warm, film-like and natural, shot for atmosphere and memory rather than technical correctness. He treats them as separate skills.
What plants does Riu grow?
Indoor plants, mostly orchids and species that cope with low light. He grows them under LED grow lights and tinkers with home automation around the setup. He's hands-on with them, enough to have diagnosed his own overwatering.
Why is Riu into space?
He's been fascinated by space since he was a kid, and can't say why. It shapes the films he rates, the books he reads and what he watches. Project Hail Mary and The Martian are both on his reading list, and he has over 120 hours in Kerbal Space Program.
Does Riu play any sport?
- Bouldering and climbing, for the problem-solving and the steady progress - Running, working toward longer distances: City2Surf, a Sydney half marathon and the Wings for Life World Run in 2026. He trains by heart-rate zone and enjoys the data. - Captainball, a netball-like sport popular in Malaysia and Singapore, which he plays because his friends do - The gym, as a general base
Who is Sylvester?
Sylvester is an Australian water dragon that visits Riu's garden. He's wild rather than a pet, but he has a name and a few photos on this board.
Where has Riu travelled?
Something over 40+ countries. Asked for a favourite he lands on Japan, Taiwan, Mexico and Malaysia, and spotted the thread himself: all four have a great cuisine and a strong living culture. Japan is the most-visited by a wide margin. Lately it has been motorbiking Vietnam and remote-reef diving across Indonesia and the Pacific.
What would Riu tell me to watch?
Andor, which he rates above the original Star Wars trilogy — "by far the best piece of Star Wars media" — for its themes on authoritarianism, which he thinks are "really relevant to us in this day and age, with rising authoritarian powers and democracy seeming to be on the decline." He describes it as a careful, slow-burn, considered drama that almost just happens to be set in Star Wars, and he's watched the first season two or three times. Rewatching is his actual quality signal: what earns it is having more themes than one viewing can hold. Worth noting he's a lifelong Star Wars fan who rates most of it badly. Being a fan of something buys it no leniency from him.
Why did Riu get into design?
It started in high-school robotics, where he found the hardest problems were usually solved by smarter design rather than more engineering. That pulled him toward digital design, where ideas can be tried and changed much faster than physical design.
What is Riu like to work with?
Direct, and easy to disagree with. - He'd rather be told he's wrong than agreed with politely, especially on anything high-stakes, and he pushes back the same way. - He wants the reasoning behind a decision, and the trade-offs talked through before anyone settles on an answer. - He gives clients choices they can actually compare: two or three real, clickable directions per page, side by side, then the strongest parts combined after feedback. - He builds what a project is missing. When a client's feedback kept arriving by email against a list of links, he built a tool for commenting on the page itself, which became the review kit. - He's dyslexic and works best from visual, structured material.
What is the koi pond?
A pixel-art koi pond Riu drew to a single canvas, at riufukazawa.com/koi-pond. It started as his first-year design programming assignment in 2022, built by hand in p5.js before AI. In 2026 he rebuilt it with AI the way he'd pictured it, with each koi an agent that wanders, flees taps and chases the food you drop. He open-sourced it as a single React component under the MIT licence, then grew it into a large board you can pan and zoom, which runs at 60 fps on his phone. Its gesture engine later fixed the lag on this site's photo board.
What's Riu's experience with the Adobe suite?
Years of it, much of it for clients: - Photoshop: campaign assets for ESPN's Australian marketing, including out-of-home posters for ESPN on Disney+, compositing AI-treated player photos into the artwork. - Lightroom Classic: more than ten years of photography, including paid event and graduation work since about 2022. - Premiere Pro and After Effects: commercial video editing, from debating-competition videos to wedding and corporate work. - Illustrator: vector work alongside the rest. These days much of his design happens in code, but the suite is still how he produces images and video.
Is Riu a front-end developer?
He can do front-end development, in React and Next.js, and he builds everything he designs. It isn't the job he's after, though. What he wants is to solve real problems through design, especially how people work with AI, and building is how he gets those designs all the way to something people can use. So a role that only turns other people's designs into pages would use half of what he does. One where he shapes the problem and builds the answer uses all of it.
What's Riu's experience with AI image and video generation?
A lot of it is client work at Now We Collide, mostly on Higgsfield. His biggest job was an image and video library for an infrastructure client whose facilities can't be photographed, which he directed end to end and took to 4K video. He also made AI player treatments for ESPN's Australian campaign assets, and built the generation pipeline into ImageGen, the agency's tool for on-brand social content. What he learned: approve a still before anything is animated, let the model make pixels while composition, timing and colour stay in code, and measure every "this looks off" so a fix can be checked. His verdict is that where filming is possible, filming is still faster, cheaper and better. - What he learned about AI video - ImageGen
Is Riu a software developer?
No, he's a design engineer, and the engineering is real. His speciality is design; the engineering is what lets him take a design all the way to working software. - Front end: React and Next.js, for client sites and for this one. - Back end for his own products: databases, image storage, caching and response times. This site's pre-written answers, its photo board and its logs all run on that. - AI harnesses: the brain has a fast mode and a deep mode that reads notes through its own tool loop, with cost limits and a privacy check. He wouldn't claim a senior engineer's depth. He does bring enough technical ability to design within real constraints and build what he designs.
Is Riu a vibe coder?
He writes most of his code with AI, and he's open about it. What separates that from vibe coding is that he designs the system and checks the result, and you can see it on this site: - the pre-written answers live in a database with a two-minute cache and a saved copy the site falls back on if the database is down; - a privacy check runs over everything public at build time and fails the deploy if a blacklisted fact appears; - he measured what each answer costs and redesigned the fast mode when it was too expensive; - the brain runs two modes, a fast one and a deep one with its own tool loop. He calls himself a design engineer: a designer whose output is working software.
Where does Riu think AI product design is going?
Towards designing what could only exist because AI exists, rather than bolting AI onto things that already work. AI at Checkout is his example: no chatbot and no visible AI, just a confirmation page that points out what a traveller didn't know to ask. He thinks the interface's job is to show what the AI is doing rather than hide it, so a person can direct it and decide whether to trust it. And as making things gets cheap, he expects judgement to be the scarce part: knowing what's worth making, and where a person has to stay in the loop. - AI and taste - AI and trust
What does Riu think about AI and taste?
That an AI agent given a person's taste drifts back to the model's own within minutes, and that it's structural, not a prompting failure. Training pulls models towards the typical answer, a person's instructions shrink as a share of a long session while the model's own output grows, and a model judging its own work prefers what's familiar to it. He learned it the hard way: he built a multi-agent build loop in 2026 and retired it in October. So he uses agents for the parts of a build meant to sit at the design system's average, and keeps the judgement that turns a brief into a direction for himself. A model trained on great designers is still an average, just a higher one. - Two kinds of taste - Is he critical about AI?
How does Riu design for accessibility?
From the start, not as a retrofit. Foodhunt, his capstone, was designed first for people with cognitive and learning disabilities, and three of the four on the team have one themselves. The team boiled their research down to three principles, navigation, communication and cognitive load, and checked every design call against them. It also meant cutting what didn't work: testers found a physical basket divider belittling, so it went. And he designs from experience: he's dyslexic, and knows how much clear, structured, visual material helps. Some places to start: - The research behind it - What testing changed
How did Riu design a visual language for AI agents?
The problem was how to signal that an AI agent is doing something, without the generic sparkle everyone reached for in 2026. His first system was a personified mark — a person glyph for agents that assist, an object glyph for automations — chosen specifically because he reverse-engineered it from how agents already appeared in the product's own screenshots rather than inventing it from nothing. That was superseded by an 8-bit pixel-agent system which wasn't his idea, but which he prototyped and pushed; the differentiator they landed on was showing agents in action, presenting work and handing it to each other, instead of posing as icons. What he originated was the governance: a rule for exactly when an agent may appear, because scattering them decoratively would teach users nothing. That rule became part of the brand documentation, and the language has outgrown the website into product identity.
What is AI at Checkout?
A university project from 2024 that uses AI with no chatbot and no visible AI at all. After you book a trip, the confirmation page uses the booking data against the user data the platform already has to tell you what you'll need, like the transport card to buy in Tokyo or how strong the UV is in Australia. The idea behind it is that people can't search for things they don't know they need to know. So the AI compares your situation with the destination and points out the gaps, instead of recommending things you might like. He started with an AI recommendation engine and dropped it, because ranking is a problem ordinary algorithms already handle better. It was quite technically impressive back in mid 2024. He had to make the model reason in four hand-built reasoning steps before ChatGPT launched its first reasoning models. Newer models have caught up and superseded the technical uniqueness of the reasoning system he built.
How does Riu show design options to a client?
As working websites rather than mockups. For each page he builds two or three directions, each a complete version with one clear idea behind it, and puts them in one place where the client can flick between them. After feedback he doesn't just pick a winner. He takes the strongest sections from each option and combines them. He also settles early where the site will finally live, like which content system, because that decides how it should be built. Once a process like this has worked once by hand, he writes it up so an AI agent can run it next time, with the checkpoints where he wants to make the call built in.
How does Riu hand a site over to a client?
He settles where the site will live at kickoff, because the platform a client will maintain it in shapes the whole build. Then he designs in code, as real clickable directions the client compares side by side, and builds the approved one on that platform. For sites a client team edits after launch, he now pushes for a headless CMS such as Sanity rather than HubSpot. From a developer's side it's much easier and faster to build on, and the sites end up with far better functionality and look much better. His earlier site for a financial services industry body went the HubSpot way: a coded theme with brand tokens, templates and editable modules, proven in the client's own portal.
How does Riu make design decisions?
On evidence, even when that means cutting something he designed and liked. Foodhunt's basket divider is the example: testing showed users found it belittling, so it went. He'll also rework something that technically works but doesn't feel right. When he gives options, the first is the most sensible choice that fits the system, the second is deliberately the most different, and there's usually one that pushes further even if it loses. That turns a review into a real choice instead of a rubber stamp. When a problem keeps coming back, he looks for the cause underneath rather than patching each case. And a clever idea gets cut the moment it reads wrong to the person it's meant for.
What is Foodhunt?
Foodhunt is a supermarket navigation app designed first for people with cognitive and learning disabilities, like dyslexia, ADHD and autism. It plans a route through the store and guides you aisle by aisle, so you don't have to work through a list of categories. It was Riu's final-year capstone at the University of Sydney, built over 13 weeks by a team of four, and he later rebuilt it on his own in React. The case study has the overview, and the brain has the full record. Some places to start: - The research - What testing changed - How he sees it now
What did user testing change in Foodhunt?
It removed one of the project's core features. The team had designed a physical basket divider, with colour-coded sections, meant to stop impulse buying and make shopping easier to manage. In the second round of A/B testing, participants found it belittling, said it took up space they needed, and didn't use it. A feature meant to reduce cognitive load was adding to it. So they dropped it and shipped the app on its own. Riu points to that as the most important moment in the project, more than the research numbers. Looking back, he thinks the research was stronger than what they made from it. The secondary research and testing were good, the survey had too few responses, and the final app was a safe outcome for how much the research had uncovered.
Tell me about a hard call Riu made on client work
Two he'd point at. An early homepage strayed from the client's approved copy deck, and while his design reasoning was defensible, he'd buried the divergence instead of flagging it — the lesson he took wasn't "don't deviate", it was never let a client discover a deviation on their own. The second: when his team's internal preference and the client's actual feedback disagreed on a homepage direction, he built the one the client had given zero criticism on rather than the one the team liked. That's the pattern generally — he'll cut something he designed and liked when the evidence says it isn't working.
What is the image generation tool Riu built?
ImageGen, a web tool at Now We Collide for making a client's social content on brand, now called Collide Content Studio. It has a library of the client's photos and clips, AI generation that creates images from the client's own styles, animates photos into short clips or improves a photo, and a four-step content maker that lays it all into brand templates and saves every export as a recipe. It began as separate copies per client that had drifted apart. Riu turned it into one codebase where each client is a brand pack deployed as its own site, built out the generation pipeline and simplified the interface. The product vision is the agency's; the build is his.
What is the review kit?
An internal tool Riu built at Now We Collide for client website reviews. It drops onto a site in progress and lets the client click anywhere and leave a comment, with no login. The part he cares about is what happens next. Claude reads the open comments and sorts each one: a small fix it makes itself, a bigger change where it suggests a few options, a question it answers, or something it flags because it isn't sure. Approved changes go onto a branch and the comment closes on its own; anything uncertain goes to a person. He designed it with the AI as one of its users. Comments are never deleted, only marked resolved, so the agent never works on the same one twice. The setup guide is written for an AI installing it, after one guessed a setting and broke a site. The full write-up is on the review kit page.
What did Riu do at SUEDE?
SUEDE is the University of Sydney's experience design society. Riu joined its industry events team in 2024 and was Industry Events Director for 2025, running a team of about four. Over that time the team ran 13 events and grew from three industry partners to nine, including Atlassian, Microsoft, Canva, Google, IBM, Deloitte Digital and Macquarie Group. He favoured events that gave students practical knowledge, like portfolio critiques and small breakout groups with designers in the discipline they cared about, over general networking nights. He's honest that not every event came off. Several fell through when partners ran out of capacity, and he postponed one alumni panel to the next year rather than run it unprepared. It went ahead in 2026, after he'd left.