Work record · Isaac Sven Manzanarez
Product designer who writes production front-end. Five years in product design. This page is the long version of my résumé: the same claims, with their denominators, methods and limits attached. It exists so I never have to shorten a number into something I would have to walk back.
A lead-qualification platform for luxury real estate. I did the research, the concept, the interface and the production code. Live with a paying client, Century 21 Heavenly in Los Cabos.
Next.js 14, TypeScript, Supabase, Tailwind, Vercel. Multi-tenant, with tenant isolation enforced at the database by row-level security rather than in application code. Scoring is a pure deterministic function and the rubric version is pinned to each lead, so a score stays explainable against the rules that produced it months later. Spanish-first, MXN pricing, LFPDPPP-compliant with versioned consent timestamping.
A desktop research browser, Electron. My own product: the design, the priorities and the specs are mine; the code is written by AI agents working from them.
v0.9.0 public beta. 115 installer downloads, 111 commits, 22 end-to-end test suites, public repository. Its own design system, including duration tokens, interaction rules and 11 animations.
Consumer iOS social product for pet owners, Toronto and New York. Product designer and founding member. Design system, onboarding, the Neighbourhoods cross-surface model, and the paid media.
Around 100,000 downloads over five years. Peak user base 13,000.
Three of the numbers below describe the same product and measure different things on different populations. Collapsing them into one claim is the error this table exists to prevent.
| Figure | What it measures | On what base | Period | Method |
|---|---|---|---|---|
| 61% → 79% | Onboarding completion | People who start onboarding | Before the 2025 redesign, and now | Instrumentation. Before and after, not a randomised split |
| 29.3% · +18pp | The same change, relative and absolute | Same | Same | 61.1% → 79.0% is 29.30% |
| 45.6% | Profile completion | All downloads, around 100,000 | Five-year average | Instrumentation |
| 13,000 | Peak user base | — | Peak, not concurrent with the download total | Instrumentation |
| >80% | Task completion inside the study | 10 study participants | The study itself | Moderated observation against a threshold set beforehand |
| $1.37 → $0.57 | Cost per acquisition | Paid media, Instagram and TikTok | Over the same period as the product work | Ad platform reporting. From the media buying, not the onboarding |
| 13.5% | Likes per session | Feed sessions | After shipping spring-based like feedback | Mixpanel Lift, 2.27 against 2.00 |
| ~60% · 70% | Design-to-code translation time, and visual inconsistency bugs | Team workflow | After the token pipeline | Internal tracking |
| 1.1M CAD | Gross profit | The first two of four venues | May 2019 – December 2023 | Books |
| 10% | Net profit margin | Same two venues | Same | Books. Industry band is 8–13%, lower in high-rent markets |
| 32% → 28% | Cost of goods sold | Venue operations | 2023 to 2026 | Books, after a pivot to a leaner model |
The finding. Pawmates onboarding had two profiles, one for the owner and one for their dog. Across ten moderated sessions, people were patient and talkative building the dog's profile — volunteering mood, curiosity, habits, the things it is obsessed with, humour, likes and dislikes, whether it photographs well — and flat building their own. The owner's profile came first in the flow.
So the problem was not the number of fields. It was which subject the fields were about. Filling in the dog profile was not data entry; it was presenting their dog to other people, and on a social product that input is the content.
Why it matters that the owner profile came first. The friction arrived before the willingness, so nobody had spent their patience when they reached the part they resisted. Fatigue does not explain it.
What a funnel alone would have supported. Completion was low, so shorten onboarding. That means cutting fields evenly across both profiles, which would have cut the one part people were motivated to finish.
What changed. The two profiles swapped order. The dog profile moved to the front, so the part people want to do earns goodwill that carries them through the personal profile, permissions and the tutorial. Four steps were resequenced behind it rather than deleted, and the location permission was deferred past the first feed view. Progressive disclosure and reserved space for error messages were applied where motivation was lowest, not uniformly.
A second check on the fatigue reading. After the change the mundane steps come later in the flow, and completion still rose. If accumulated effort were the cause, pushing that work further back would have made it worse.
Protocol. Ten participants, mixed ages and tech comfort, recruited through local pet groups, Instagram and known owners. The sample deliberately included people who do not own pets, because one of the four tasks asked what the app is for and a target user already knows. Moderated, remote or in person, 30 to 40 minutes. Think-aloud, screen and audio recording, one facilitator and a note-taker. Four tasks: start onboarding, build the dog profile, say in your own words what the app is for, find another dog nearby. Measured on task success, time on task for profile creation, errors and backtracks, think-aloud comments, and a threshold of more than 80% reaching a first content view without getting lost.
Think-aloud is why the study produced a finding and not only a number. The asymmetry showed up in what people said while working, not in what they completed; a click-tracked test would have shown the same drop-off with no reason attached. Later iterations ran a cut-down version — five users, three tasks — small enough to run every cycle.
A 2026 pass doubled down on the same effect rather than revisiting it: a progression bar, bigger completion buttons, and tighter spacing so the progression feels compact. The sequencing decision held through a full rebuild, and completion kept improving.
On contract in 2026 as product designer and PM, working with the founder as decision owner, two engineers and a content team.
Sixteen locked or recommended decisions, each with a named owner, a date and a status, carried forward in every spec. A locked decision was not reopened unless new information arrived. Open questions always had an owner and a default already running in the prototype, so the founder could react to a working state instead of a description.
A weekly founder sync covering every track, with a numbered list of decisions needed that day. A per-milestone session log recording what changed, what was decided, what was still open and what had gone stale. A per-build engineering review documenting what existed first, then issues ordered by priority, written as observations without assumptions about the code.
Each recorded as before, after, trigger and the trade-off accepted. Five proposed ad products became one. A campaign tool became a label with four interactions, because there was no backend data to make tracking meaningful. A breadcrumb with nested tabs became a single scope selector, keeping the tiers and removing only the nesting, with a feed cascade so the empty-feed problem did not return.
Push notifications were the missing reason to return, so building them first looked obvious. I recommended against it, in favour of an SMS and email rail carrying the same four message jobs, so a density pilot could start immediately instead of waiting on an engineering build. I named the risk in the same document: a null pilot result could be caused by the weaker channel rather than by a missing density effect, so the read would have to account for that.
People will do effortful input to present their dog and resist the same effort about themselves. Neighbourhood member counts are not printed, because a neighbourhood with three members looks dead if you print the number, so activity shows as qualitative labels. Like has immediate spring feedback; unlike is deliberately not animated, because it is not a behaviour the product rewards.
The beta has eleven tracked metrics tied to four questions — does it work, does it help, do people come back, who is it for — with the source named for each and targets set in advance and labelled as first guesses to reset after the first review.
Decision rules are checked in order and stop at the first that applies, so the next cycle's scope is decided by the scorecard rather than by preference. Any open blocking bug means stability work only. Fewer than five survey responses means the next cycle works on reach. Suggestions rated low means the keyword engine. Testers trying it once and stopping means setup and first run. The beta ends only after two consecutive cycles with no open blocking bugs on all three platforms, at least 40% of fifteen or more respondents saying they would be very disappointed without it, 30% still using it after seven days, and most saying source-finding is faster.
Current state, honestly: two survey responses against 115 downloads. Update checks per download have risen from 2.0 to 2.3, which is the only usage signal available because NODE sends no data. Both respondents name the same feature as the one they would keep, and both name privacy as the reason they have not switched — neither mentions research speed, which is what the success criteria are built around. At two responses that is a hint, not a finding. If it holds, the product is working and the positioning is pointed at the wrong benefit.
I founded CASA Hospitality Group in 2019 and still run it. Four venues over its life. The first two returned 1.1M CAD gross profit between May 2019 and December 2023, at a 10% net profit margin against an industry band of 8 to 13%, lower in high-rent markets. A pivot to a leaner operation — less overhead, higher post-COGS margin per dish — took cost of goods from 32% to 28%, and that is what made the third venue, Casa Mezcal, attractive to the investors who acquired it. It was handed over. Rumbar Mezcal in Connecticut operates today.
I ran Toast POS and OpenTable reporting, built the FOH and BOH procedures, inventory and training playbooks, and held labour under 30%. The reason this is on a design record: it is where I learned what operational software gets wrong. A mistake is caught during service, with the guest in front of you, fixed informally in thirty seconds, and the software never learns it happened.
Cal.com's design system, built on headless primitives. Dialog and Toast now respect prefers-reduced-motion. I measured every running transition in Chromium with a harness I built for it, then put the guard at the primitive layer so it holds for every consumer rather than at each call site. Under reduced motion both keep their fades and stop moving.
Dark-mode tokens were identical to light mode. Moved the palette into the shared package and fixed a contrast failure: 2.57:1 to 7.29:1.
On the first, I showed that the RTL suite still passed with every directional class stripped — the behaviour the change rested on was not covered by a test.
BDes (Hons) Industrial Design, OCAD University, 2018–2023. Product Management Certificate, BrainStation. English, Spanish, conversational French.