0 mSurface
Josue Perez
Product designer. I work on systems where AI is a material rather than a feature — which mostly means deciding what a machine should not do, and making the seams legible to the person holding it.
Selected work
Compass
A personal ecosystem that could not notice its own failures — rebuilt as a two-minute morning glance. Shipped; in daily use; silent failures no longer silent.
ReadThroughline
People navigating institutional systems re-tell their story at every doorway. A continuity layer, designed end-to-end; bilingual prototype shipped, honestly unpiloted.
Read12 mShallow water · still full colour
How I find out what is true
Everything above this line is a claim. This layer is the method that has to hold it up.
Every claim at one depth is evidenced at the next one down. If a layer cannot be defended by the layer beneath it, it comes off the site.
- Research
- Contextual interviews and observation over surveys. I care most about the gap between what people say they do and what the logs say they did — that gap is usually where the product actually is.
- Interaction design
- Designing the state machine before the screens. Most interface bugs are modelling bugs that arrived late.
- Behavioural design
- Used to reduce error and effort, not to manufacture engagement. I will argue about this one.
- Evaluation
- Usability testing with tasks that can fail. A session where nobody gets stuck told you nothing.
30 mReef · red is gone
Where the craft compounds into a system
Individual good decisions do not add up on their own. These are the projects where the work was the structure connecting them.
Compass
The constraint was attention, not information. The contested decision: a hard three-item cap and one next move — completeness lost the argument. Shipped as an offline-capable PWA; the number that moved is silent failures lasting over 48 hours: zero since launch.
Full case studyThroughline
A continuity layer for immigration, healthcare, and benefits journeys. What did not work: the first information architecture organized by institution — users think in life events, not agencies. Rebuilt around moments; the prototype is bilingual because its users are.
Full case study- Systems thinking
- Looking for the feedback loop rather than the feature. Most "engagement problems" are a loop somebody built on purpose and then forgot about.
- Journey mapping
- Only useful when it includes the parts outside your product. The support call, the spreadsheet, the colleague who knows the workaround.
- Automation
- The design question is never "can this be automated" — it is "who is accountable when it is wrong, and can they see enough to be".
150 mBlue water · edge of the sunlight zone
Bets, stated as bets
Work without a category yet. I would rather show you what I am wrong about early than only show you finished things.
- AI-native workflows
- Not "AI features bolted to an existing flow". Asking what the flow would be if the model had been there from the first sketch — and usually finding it has fewer steps and more checkpoints.
- Agent ecosystems
- Several agents, one accountable human. The unsolved design problem is legibility: how someone stays responsible for a process they did not watch.
- PerX Labs
- The lab exists to ship working versions of designed things. The bet is wrong if a year of replicas changes no conversation — measured against interviews and engagements, not traffic.
- Safety-critical & maritime UX
- The bet: operator UX — trust calibration, degraded-mode decisions, alarm design — is one skill market across maritime, energy, and autonomy, and the water is where I train it for real. The evidence is regulatory, not romantic: S-100 navigation, the MASS autonomy code, and EU emissions tooling all land 2026–2030, all needing interfaces for people who are cold, wet, gloved, and glancing. I hold PADI Rescue Diver; STCW Basic Safety is the next rung. Wrong if, by 2028, none of it has produced a single paid engagement — measured in invoices, not enthusiasm.
600 mTwilight zone · no sunlight, some light
Research that is not product
Things I build to think with. None of this has a customer. That is the point of it.
Decision Depth
An open-water decision simulator — scuba, freediving, snorkelling, swimming — built on real decompression, gas and hypoxia models, with emergency drills drawn from published procedure. Its most interesting output is a list of its own limits.
Read the write-up · Open the simulatorInterfaces for moving environments
Glanceable state over dense state. Shape and position over color. Confirmation you can feel. The full argument lives in the Decision Depth write-up — and in the water it came from.
NotesCertifications, with limits attached
PADI Rescue Diver (2022) — managing panicked people in the water; not an instructor. Advanced Open Water (2021) — recreational to 30 m; not technical. EFR first aid (2021) — lapsed, listed anyway: a ledger that hides lapses isn’t a ledger. USCG Auxiliary & FEMA ICS — SAR, comms, incident command; training, not licensure. Next rung: STCW Basic Safety, Barcelona.
Why a designer trains like an operator4000 mAbyssal zone · no light at all
Why any of it matters
The bottom of this site is not a conclusion. It is the set of commitments that the five layers above are downstream of.
- Human-centred AI
- The phrase is nearly meaningless now, so here is the operational version: the person must be able to find out what the system did, and disagree with it, without asking permission.
- Anthropology
- Every workflow is a culture with a technology attached. Redesigning the tool without reading the culture is how good products get quietly abandoned.
- Complex systems
- Most failures are not component failures. They are interaction failures between components that each worked correctly.
- Opportunity design
- Deciding what is worth building is a design act, and it is the one with the largest coefficient. Everything downstream of a bad "why" is expensive.