# Carlos Montoya — public launch context Canonical site: https://losmontoya.com Title: Principal Product Designer & AI Engineer ## Summary Principal Product Designer & AI Engineer. I turn ambiguous product questions into working software, then build the evaluation systems that make judgment observable, debuggable, and testable. My current independent practice combines product design, native and web prototyping, design systems, and AI-native workflows. Sole inventor on a patent filing in AI context engineering. # Work Index Home is a curated selection. The fuller portfolio archive starts at https://losmontoya.com/work and is organized into five browsable categories: - Interfaces: https://losmontoya.com/work/interfaces - Prototypes: https://losmontoya.com/work/prototypes - Brand: https://losmontoya.com/work/brand - Ideas: https://losmontoya.com/work/ideas - Photos: https://losmontoya.com/work/photos # Hobbies URL: https://losmontoya.com/hobbies Photography, old lenses, and the life around the practice. # Flagship case studies ## 48 Hours to Native URL: https://losmontoya.com/work/48-hours-to-native Headline: Some version on mobile. Bet: Leadership could not judge the future of Dropbox mobile from static screens. The interaction had to run natively. Impact: A research-backed mobile thesis became a native SwiftUI app in roughly two days, giving leadership and engineering something real to judge. Evidence note: Demo data by design. State, motion, and navigation are native. ## Situation Mobile was the last and vaguest line in an executive brief about the future of AI at Dropbox. I started with research, not screens: foundational customer studies, the One Dash strategy, mobile habits, fragmented project context, and the action-and-approval model behind CoWork. The synthesis produced a product role. Mobile should not be a compressed web app. It should be the companion where a person catches up, retrieves quickly, approves agent work, and dispatches the next action. In a March 5 journey review, principal researcher Mahsino sharpened the thesis: explain why an intelligent card appears, treat Calendar as the morning-orientation competitor, keep Library as a familiar safety blanket, emphasize high-value field and retrieval contexts, and surface passive oversight without making managers ask for status. ## Making The direction moved from Pencil wires to high-fidelity HTML using real content and tokens. That fidelity did its job by failing in public: during the March 13 cross-platform review, stakeholders said the polished mobile screens felt like a different product. The custom glass language had drifted from the web story. I accepted the critique and made two same-evening calls: converge on the same Dropbox design language and content across platforms, and reserve native iOS behavior for what was genuinely platform-specific. I configured the environment, translated the design system into SwiftUI, and began the native pivot that night. Over roughly two days, I built real SwiftUI navigation and state, one Search and Ask surface, a minimizing bottom capsule, preview with comments and actions, and a CoWork flow that drafted and sent an update. Deterministic demo data kept the review path stable; the navigation, keyboard, sheets, scroll physics, and transitions remained native. The opening loop shows the answer-and-sources path. This second path tests a materially different question: whether Search can move from broad retrieval into specific typeahead without losing the surrounding context. Every public capture comes from the installed native app; there is no browser port or public runtime. ## Outcome In a two-week executive sprint, I turned an ambiguous mobile brief into a research-backed companion strategy, then built the key paths in native SwiftUI over roughly two days. Leadership could judge the direction in hand; iOS engineers asked for the repository, and the prototype circulated internally.

THE AFTERLIFE

The iOS engineers asked for the repository, then the mobile experience team joined the review. The prebuilt simulator app circulated internally, and I presented the direction at the Q2 read-in.

The strategic output was larger than a polished prototype. I built the decision system around it: research synthesis, traceable hypotheses, cross-platform critique, native architecture, stakeholder convergence, and an implementation artifact engineering could inspect. ## Reflection Native was the shortest honest route because the remaining unknowns were behavioral. A static screen could describe Search collapsing under a thumb or an action staying attached to a preview. It could not make interruption, reachability, keyboard timing, sheets, and state transitions judgeable in hand. The SwiftUI app remains local capture provenance only. The evidence on this page is truthful native footage, not a downloadable app, hosted demo, reconstructed web version, or production-service claim. ## Reimagined Onboarding URL: https://losmontoya.com/work/reimagined-onboarding Headline: A zero behind it. Bet: Demonstrate value before asking for work: the system observes and proposes, then the human approves. Impact: Multiple coded directions and two rapid research rounds converged on a trust model: show value, ask permission, keep every change inspectable. Evidence note: Working prototype with fixture data. ## Situation The brief asked why software onboarding could not feel like the first minutes of a great game: learning by doing, not reading a tour. The real constraint was tempo. People arrived to get work done and were open to help only briefly, so a capable product had to prove value without blocking the work. I turned that tension into four different theories of activation: a cinematic foreground takeover, a smart cursor working inside the product, a conversational experience shaped by intent, and template-driven first success. These were not visual skins. They tested whether belief came from attention, demonstration, conversation, or a seeded win. ## Making Round 1 included six participants. The clearest signal was not that one concept won. People wanted AI to do or find something useful before asking for more work, and risky actions needed approval, reversibility, and a visible trail. The foreground takeover exposed the wrong tempo: it was slow and obstructive. The part worth keeping was smaller—the self-paced, bite-size tutorial behavior that could teach one moment, tuck away, and resume later. The smart cursor produced the stronger unique signal because it demonstrated capability in the real product, then made its active or parked state visible. The prototype changed between sessions. A Round 1 synthesis in the morning became five coded changes before the next research session: fewer connector prompts, clearer object visibility, explicit active and parked controls, visible controls instead of memorized shortcuts, and copy centered on useful work. The operating model was session → code → session, often inside one day. ## Outcome The exploration expanded to eleven concepts, then converged to four directions on one architectural spine: enter the real product, let one surface take focus temporarily, show a bounded success, collapse back to the default state, and let contextual assistance handle the long tail. The observational rename flow became the convergence proof. Dash waits for a person to rename one clip, recognizes the pattern, proposes extending it, opens review, and changes nothing until approval. The opening loop shows that payoff; it is the end of the decision history, not the whole case. ## Reflection Across two rapid research rounds, I built multiple coded onboarding directions and changed the prototype between sessions. The work moved the team away from foreground takeover toward bite-size learning and an ambient trust model: show useful work, ask before acting, and leave a visible trail. The senior-principal artifact was not a favorite mock. It was the decision system connecting research synthesis, coded hypotheses, between-session iteration, product architecture, stakeholder convergence, and an engineering-ready trust model. The footage on this page comes from working prototypes with fixture data. It demonstrates interaction and decision logic, not a production service or achieved activation lift. # Ideas ## Pantone, 1963 URL: https://losmontoya.com/work/pantone-1963 Artifact: The authored 11-swatch roster runs yellow first, then light to dark. It is not spectral order. ## The thesis Pantone's breakthrough was not a color book. It was a shared language that let a fragmented supply chain coordinate. A formula-backed number connected designers, printers, and ink makers without requiring every handoff to depend on one person's eye. The business lesson is the shift from translator to reference. Pantone turned internal ink-room knowledge into a standard that other companies could distribute, specify, and build around. ## The artifact I researched, wrote, and designed an annotated business-history case. The compact visual layer uses the case's authored 11-swatch roster, ordered yellow first and then light to dark. It is not spectral order. The sequence preserves the ink-room logic documented in the close read. Corrections stay visible as marginalia. Founder lore, archival evidence, and inference do not silently collapse into one voice. Generated period imagery is never treated as evidence. ## What AI did Claude accelerated public-source research and drafting. I set the thesis, designed the artifact, and kept unresolved threads visibly unresolved. ## Taste, Versioned URL: https://losmontoya.com/work/taste-versioned Artifact: The Brand patch worked on its target and regressed Interface. The repeat confirmed both effects, so the unscoped patch was held instead of shipped. ## The corpus I made a 165-file design corpus legible enough to curate, then made evaluation failures inspectable. The coverage ledger accounts for every file: 124 yielded candidates, while the rest were explicitly classified instead of disappearing into an unknown bucket. The complete crawl surfaced 21,678 candidate nodes. The point was not to automate the final selection. It was to make a large archive inspectable enough that I could spend judgment on the work that earned it. ## The evaluation The first review run produced 592 tiles. A separate 116-tile gate carries my own labels and written rationales, so a judge can be measured against a durable reference instead of trusted because its output sounds confident. That distinction matters. One experiment exposed a missing Brand rule in the live harness. Restoring it reduced Brand keeper false-cuts from 30 to 13, but increased Interface false-cuts from 14 to 22; the repeat landed at 13 and 21. The unscoped patch was held instead of shipped. The failure became an inspectable system defect and a narrower next test. ## What AI did Deterministic filters and local models handle volume. Frontier judges are measured against my labeled rationale corpus, not trusted by default. I designed the pipeline, rubric, evals, and final human gate. The system can narrow and challenge. The outward choice remains mine. ## Podcast Brain URL: https://losmontoya.com/work/podcast-brain Artifact: The source catalog gives every downstream result a stable episode, date, and runtime to return to before it becomes reusable knowledge. ## The archive I turned 140 episodes and 75.2 hours of conversation into a searchable, correctable record. The archive makes years of accumulated thinking usable without pretending that a search result is already a fact. The durable unit is not a generated summary. It is a result that can travel back to an exact source, survive correction, and keep its uncertainty attached. ## The retrieval rule Search produces leads. An exact-source pass decides what graduates into reusable knowledge. That boundary protects the record from a fluent model collapsing wording, attribution, and interpretation into one confident answer. The source index shown here is the durable first layer: stable episode order, titles, dates, and runtimes. Because speaker attribution can still fail at the edges, any result that matters returns to its exact source before it graduates. ## What AI did Local transcription, diarization, and indexing make the corpus searchable. An exact-source pass decides what becomes reusable knowledge. I designed the archive, retrieval model, and the rule for when a result graduates to truth. ## LMDC Studio URL: https://losmontoya.com/work/lmdc-studio Artifact: Each seat records why its incumbent holds—and where it does not. The board stays contestable because a measured failure can change the routing. ## The instrument The lab is becoming a measured routing system. Repetitive extraction and skimming run locally where the lane has evidence. Higher-stakes judgment stays with frontier models and me. Measured failures change the routing. The split is economic, not ideological, and it is not absolute. A seat is provisional: it holds while it meets the test for its own job, then moves when a measured failure says it should. ## The operating boundary The lab serves two durable corpora: a design archive and a podcast archive. The seat ledger keeps three things together: the job, the evidence for the current choice, and the acceptance test that could change it. That makes uncertainty operational. Enrichment remains open because its carried throughput claim has not been measured on the box. Speech-to-text remains open until it survives a bake-off on the actual podcast audio. An open row is a work queue, not a failure of confidence. ## What AI did Local models do repetitive extraction and skimming where the lane has earned that work. Frontier models and I handle higher-stakes judgment. Measured failures change the routing. I designed the evals, operating boundaries, and final outward gate. # Experience ### Principal, Independent Design Practice, Los Montoya Design Co May 2026 - present - Built a taste-evaluation pipeline that accounted for 165 Figma files; 124 yielded 21,678 candidate nodes, 592 first-run review tiles, and a separate 116-tile labeled gate. - Benched the calibrated judge against frontier models and kept the disagreement log as the calibration asset. The rubric is versioned like code because it is treated like code. - Rebuilt and shipped production marketing sites for select clients on Next.js and Vercel, migrating them off Webflow for speed and editability. ### Principal Designer, Dropbox Nov 2022 - May 2026 - Built a native SwiftUI prototype in roughly two days because static prototypes could not prove the state and motion model. Leadership reviewed it hands-on, iOS engineers asked for the repository, and the app circulated internally. The public evidence uses deterministic demo data by design. - Built four coded onboarding directions that made an AI-led first-run thesis testable and converged on a clear trust model: the system proposes; the human approves. - Set the bar for what prototyping looks like at Dropbox, moving high-stakes design review from slides to working software. - Mentored and coached 25+ designers from AI trained to AI native through open working sessions, critique loops, and agent-assisted design practice. - Sole inventor on a patent filing in AI context engineering. - Developed principles for AI and workflow systems, covering how users move from search and reading into action, organization, transformation, routing, and automation. - Regular interviewer on the design hiring loop, running Making and Thinking deep dives and portfolio reviews. ### Founding Design Lead (fractional), then Advisor, Hamster · Wheel Go Fast, Inc. Mar 2025 - Oct 2025 - Helped define and ship an AI-native product workspace for teams building with agents, from brand identity through MVP product design, design system, and marketing surface. - Set up agentic design operations and principles for a team working alongside AI from day one. - Company shipped to early access; I remain an advisor. ### Design Director, Principal Designer, Twenty · Healthy Together Jul 2019 - Jun 2022 - Designed the first end-to-end iOS and Android contact-tracing solution for the State of Utah, from concept to launched app in 28 days during the COVID-19 pandemic. - Supported a product with a 4.9 App Store rating and 220k+ reviews across iOS and Android, on a platform now serving 5M+ people. Second state in the nation with a COVID app, first to launch on iOS and Android on day one, and the first state app using Bluetooth proximity tracing. - Led a team of four senior designers, hired as the team grew, and merged two design organizations after an acquisition. ### Staff Designer, Creative Director, Juxta Labs · Mappen Aug 2017 - Jul 2019 - Took consumer social products from conception to production and matured design capability across the product team. ### Senior Web Designer, Atlassian Apr 2015 - Aug 2017 - Led creation of web design platform systems for marketing at scale. ### Product Designer, Pinger Jan 2013 - Apr 2015 - Led product design across iOS, Android, Windows Phone, desktop, and web for Pinger, TextFree, Sideline, and more. # Provenance and claim limits - 48 Hours to Native footage is recorded from the real native SwiftUI runtime with deterministic demo data. No public app runtime, TestFlight, download, or browser port exists. - Reimagined Onboarding footage is recorded from the coded prototype with fixture data. - Taste corpus: 165 files accounted for; 124 yielded candidates; 21,678 candidate nodes; 592 run-one review tiles. The 116-tile labeled gate is a separate evaluation corpus. - Claims distinguish working evidence from shipped product impact.