Available for select product teams
Louie Sakoda
The latest model is an AI product designer who can think through the system, design the experience, and build the working product.
*Add to cart for pricing
Description
Most recently, I built Offboard, an AI-native career-transition platform using React, TypeScript, Supabase, Deno edge functions, and multi-model LLM workflows. Before that, I spent nine years at CK-12 designing AI tutoring and learning products used at scale.
Product Specs
- Solo builder of a full-stack AI product using React, TypeScript, Supabase, Deno edge functions, and LLM-assisted development
- 9 years designing K-12 learning products, including AI tutoring and teacher-facing tools used at scale
- Strongest in ambiguous product spaces where AI workflows, user trust, and hands-on prototyping all matter
- Experienced across product design, UX architecture, front-end implementation, design systems, and human-in-the-loop AI
Skills
Ask Louie AI
Skills
Book Some Time
Work
Check out some of my work below
Offboard - Jobseeker OS Application
I designed and built OffboardOS, an AI-native job-search operating system that helps laid-off tech workers move from overwhelm to action with guided workflows, resume tools, interview prep, and personalized AI support.
Designing Offboard, a Full-Stack AI Job Search OS
Designing Offboard, a Full-Stack AI Job Search OS
Offboard started from a pretty simple observation: Job seekers do not need another disconnected tool. They need one place where the work around a role can stay connected.
I personally designed and built Offboard as an AI-native career-transition platform for people navigating layoffs and job searches. The product helps users turn messy inputs like job URLs, resumes, company research, and interview notes into structured next steps.
The main workflow is the Job Packet. A user can paste in a job URL and Offboard builds a role-centered workspace around it. From there, the system can parse the job, check for ghost-job signals, save the application, generate company research, analyze fit, tailor a resume, draft a cover letter, and move the packet into review.
My role covered the full product surface, from product strategy and UX architecture to AI workflow design, front-end implementation, Supabase architecture, QA, and iteration.
Built with React, TypeScript, Vite, Tailwind, shadcn/ui, Supabase, Postgres/RLS, Deno edge functions, Stripe, Resend, OpenAI, Anthropic, ElevenLabs, Firecrawl, Claude Code, Codex, and Lovable.
Project: Offboard
Product type: AI-native career transition platform
Audience: Laid-off and actively searching professionals, with adjacent employer-sponsored outplacement workflows
My role: Solo product designer and technical builder
Work: Product strategy, UX architecture, AI workflow design, interface design, front-end implementation, Supabase architecture, QA, and iteration
Stack: React, TypeScript, Supabase, Deno edge functions, OpenAI, Anthropic, ElevenLabs, Firecrawl, Claude Code, Codex
Thesis: A job search should not be a pile of disconnected tools. Every role should become a connected workspace with context, artifacts, decisions, and next steps.
The Problem
Job search creates two kinds of pressure at once.
The first is practical. People have to track roles, read job descriptions, judge whether a posting is worth their time, tailor materials, prepare for interviews, follow up with contacts, and remember what happened across dozens of moving pieces.
The second is emotional. They are often doing that work while stressed, uncertain, and isolated.
Most job-search tools only solve one slice of the workflow. A resume builder helps with a resume. A tracker stores applications. A job board shows roles. Notes apps hold interview prep. Chatbots can generate advice.
The problem is that the job seeker still has to hold the whole system together.
That fragmentation creates a hidden tax. Users have to remember where everything lives, decide what to do next, and translate context from one tool to another while they are already under pressure.
The design challenge was not just to add AI to job search. It was to make the search feel less scattered.
Product Thesis
Offboard is designed around a simple idea.
Every meaningful job-search action should compound into better context for the next one.
If a user uploads a resume, that context should improve role match, resume tailoring, and interview prep. If they paste a job URL, the system should be able to parse the role, check whether the posting looks risky, save the application, generate company research, create tailored materials, and keep everything attached to the same opportunity.
The product centers on four connected layers:
- The user context: Profile, resume, goals, preferences, network, conversations, and allowed personalization signals
- The role/company context: Job posting, company research, ghost-job signals, role match, and application status
- The generated artifacts: Tailored resume, cover letter, interview prep, notes, and follow-up actions
- The guidance layer: LUMO, a context-aware AI assistant that can suggest actions, explain next steps, and ask for confirmation before changing user data
The goal was to move the experience from “use five tools and stitch the results together” to “start from one opportunity and build the right packet of support around it.”
My Role
I personally designed and built the product experience end to end.
My work included product strategy, UX architecture, AI workflow design, interface design, front-end implementation, Supabase-backed product logic, Edge Functions, QA, and iteration.
Key contributions included:
- Defined the product strategy and operating-system framing
- Mapped the job-search workflow from setup to application execution to interview prep
- Designed the core UX architecture across dashboard, applications, job packets, resume tailoring, ghost-job checks, LUMO, and interview prep
- Designed AI workflows around user control, confirmation-first actions, and risk gates
- Built front-end flows and integrated them with Supabase-backed product logic and Edge Functions
- Shaped the brand and interface direction around calm, practical, emotionally aware execution
- Used LLM-assisted development to prototype, validate, and iterate across a broad product surface
Key Workflow 1: Job Packet as the role-centered workspace
The Job Packet is the clearest expression of the product thesis.
A user can paste a job URL or job description, and Offboard builds a role-specific workspace around it. Instead of creating disconnected outputs, the system keeps the role, company, application status, generated materials, and next steps tied together.
The workflow can:
- Parse the job
- Run a ghost-job check
- Pause at a risk gate if the posting looks risky
- Save or reuse the application
- Generate company intel
- Analyze role match
- Path-to-a-Person search
- Find or request a resume
- Create a tailored resume session
- Draft a cover letter
- Move the packet into review
The important design decision was the risk gate.
If a posting looks questionable, the system pauses before asking the user to spend more time tailoring materials. That moment matters because job seekers do not just need more automation. They need help deciding where their energy is worth spending.
The Job Packet also gives the product a durable organizing object. Each role becomes a workspace with its own context, artifacts, and status instead of another loose item in a tracker.
Key Workflow 2: LUMO as a context-aware guidance layer
LUMO is the AI layer that sits across the product.
The goal was not to create a generic chat surface. LUMO needed to understand the user’s real search context, including profile, applications, interviews, network, reflections, saved chat memory, and any personalization signals the user allows.
LUMO can answer questions, suggest relevant tools, and propose actions like creating or updating an application or saving a contact. For actions that modify user data, the interface uses confirmation-first cards so the user can see exactly what will happen before approving it.
The key design decision was to keep context and control central to the AI experience.
The more the system understands about the user’s search, the more important it becomes to show what it knows, ask before changing data, and keep suggested actions grounded in existing workflows.
LUMO should feel less like a chatbot bolted onto the product and more like a guidance layer that helps the user move through the system with less friction.
Design Principles
1. Calm before clever
The product is for people who may be anxious, tired, or overwhelmed. The interface needs to feel clear and steady. That shaped the language, density, workflow rhythm, short next steps, visible status, quiet hierarchy, and practical calls to action.
2. AI should preserve agency
The system can automate analysis, drafting, and organization, but it should not make irreversible decisions on the user’s behalf. Confirmation cards, review flows, and risk gates make automation feel more trustworthy.
3. The role is the anchor
Most job-search tools produce disconnected outputs. Offboard’s workflows keep artifacts tied to the same role and company context. This makes the product feel cumulative instead of transactional.
4. Context should compound
Every completed action should make the next action easier. A resume improves tailoring. A saved role improves interview prep. A ghost-job check informs whether to invest more time. LUMO becomes more useful as the workspace fills in.
5. Progress should feel operational, not performative
The dashboard uses checklist progress, pipeline status, packets, and upcoming events. These are practical signals, not artificial achievement loops.
Design decision: I made context and control central to the AI UX. The stronger the system gets at understanding the user’s search, the more important it becomes to show what it knows, ask before changing data, and keep suggested actions grounded in existing workflows.
Outcome
Offboard is now a working AI-native career-transition product with shipped surfaces for dashboard guidance, applications, resume tailoring, ghost-job analysis, Job Packets, LUMO guidance, interview prep, documents, network, community, financial runway, and employer-sponsored support.
What is publishable now:
- The product is more than a concept. It is a real React and Supabase application with working routed workflows.
- The Job Packet pipeline connects parsing, validation, application creation, company research, role match, resume tailoring, and cover letter generation.
- LUMO is designed as a context-aware guidance layer with confirmation-first actions.
- The core design strategy is grounded in reducing scattered job-search work into a role-centered operating system.
What I would improve next:
- Add stronger usage analytics around which workflow steps create the most momentum
- Tighten the handoff between LUMO suggestions and deeper product surfaces
- Add clearer empty states and recovery states for partial packet generation
- Continue improving the trust model around AI-generated recommendations and user-controlled edits
AI Legal Consultant
'We Are Not Lawyers' is an AI-native platform that turns confusing legal problems into clear, step-by-step actions through guided workflows, documents, and optional attorney handoff.
Lumo - Jobseeker Companion
Lorem ipsum dolor sit amet, consectetur adipiscing elit.
Project Overview
Flexi is an AI-powered digital tutor designed for CK-12 Foundation’s educational platform, offering personalized, on-demand academic support for K-12 students. By delivering explanations, adaptive practice, and encouragement, Flexi addresses critical learning gaps, particularly in remote or hybrid classroom settings.
- My Role: Senior Product Designer (User Research, UX/UI Design, Prototyping)
- Timeline: January 2024 – September 2024 (Beta), Public launch Q1 2025
- Team: 1 PM, 3 ML Engineers, 2 Curriculum Specialists, 1 Front-End Developer
Problem & Opportunity
The shift to remote learning revealed significant gaps in personalized student support:
- Only 50% of students could consistently focus during remote lessons, with just 41% feeling motivated (YouthTruth, 2022).
- Teachers experienced burnout from repetitive student queries, averaging 54-hour workweeks, with less than half dedicated to teaching (Education Week, 2022).
- Effective 1:1 tutoring remains prohibitively expensive despite proven efficacy, delivering +0.20 to 0.23 standard deviation improvements in math (NBER, 2022).
Opportunity: Build a scalable AI tutoring tool integrated into existing classroom workflows, offering affordable, trustworthy, always-available learning support.
User Research & Methodology
We employed a variety of rigorous research methodologies to ensure our solutions directly addressed real user needs:
- Student Interviews (Grades 6-11, n=18): Conducted qualitative 1-on-1 video interviews to deeply understand students' emotional responses, frustrations, and expectations when seeking help.
- Teacher Diary Studies (n=12, two-week duration): Teachers documented daily interactions, highlighting repetitive support tasks and workload impacts.
- Competitive Analysis: Systematic assessment of 5 leading AI tutoring products, evaluating usability, transparency, teacher integration, and trustworthiness.
Key findings:
- 67% of students reported frustration without immediate help.
- Teachers spent ~7 hours weekly addressing routine clarifications.
- Competitors lacked transparency, teacher oversight, and trust-building features.
Design & Prototyping
Based on research, we developed a conversational, supportive user interface emphasizing transparency and accessibility:
- Confidence Indicators: Real-time display of AI confidence levels, increasing trust and clarity.
- Interactive Learning Loop: Answer → Micro-quiz → Stretch-prompt, promoting deeper learning and engagement.
- Teacher Dashboard: Real-time analytics highlighting common student misconceptions, streamlining teacher interventions.
Through iterative Figma prototyping and usability testing (n=31), we achieved:
- 22% faster task completion compared to traditional worksheets.
- 92% student satisfaction rate, measured by willingness to use the tool again.
AI Integration & Design Decisions
Strategic AI design choices were grounded in evidence and aligned with user needs:
- Transparency & Trust: Displayed model confidence and source citations, crucial for sustained trust (Meta-review of Intelligent Tutoring Systems, 2023).
- Promoting Metacognition: Implemented "think-aloud" checkboxes, shown to boost engagement and retention during remote learning.
- Teacher Empowerment: Added teacher-controlled toggles (e.g., "Pause Flexi," "Rephrase Response"), reducing teacher workload and enhancing control.
Pilot Results & Impact
Flexi’s pilot launch demonstrated significant improvements across key metrics:
- User Engagement: Weekly active users hit 61% of the target cohort within eight weeks.
- Session Duration: Average session length more than doubled, from 5m 22s to 10m 48s.
- Learning Gains: Accuracy on follow-up tasks improved significantly, rising from 48% to 72% correct responses.
- Teacher Workload: Support requests decreased by 33% per student per term.
Flexi’s initial success projects substantial long-term educational benefits:
- Estimated to save approximately 1.9 teacher-hours per class weekly, equating to nearly $2,400 saved per teacher each semester.
- Achieved an estimated learning improvement of +0.18 standard deviations, comparable to traditional human tutoring at a fraction of the cost.
Post-launch, Flexi recorded over 500,000 student sessions within 90 days, solidifying its value to CK-12’s expansive user base.
Reflection & Future Vision
The Flexi project reinforced critical lessons in designing responsible and effective AI educational tools:
- Successful: Confidence indicators, micro-quizzes, and teacher empowerment controls resonated positively.
- Areas to Improve: Early avatar designs felt overly juvenile to older students; future iterations will embrace a more universally appealing aesthetic.
Looking ahead, we plan to expand Flexi's capabilities:
- Multilingual Support: Broadening global accessibility.
- Adaptive Content: Adjusting complexity based on individual reading levels.
- Enhanced Analytics: Providing district-wide insights through API integrations.
Flexi: Designing an AI Tutor That 265 Million Learners Could Trust
How we made an AI tutor safe, honest, and actually useful for a 12-year-old, without cutting the teacher out of the loop.
Flexi: Designing an AI Tutor That 265 Million Learners Could Trust
How we made an AI tutor safe, honest, and useful for K-12 students, and what an independent classroom study taught us about where it worked and where it didn't.
- My Role: Lead UX Designer (research, UX/UI, prototyping)
- Team: 1 PM, 3 ML engineers, full content and curriculum specialists, 4 front-end devs, 2 UX designers, 1 QA. Remote Team: illustrators, 3d animators, front and backend engineers
- Timeline: Jan 2024 – Sep 2024 beta, public launch Q1 2025
- Platform & Users: CK-12 Foundation - 265M+ users served, 2.05B+ questions answered
The Setup
CK-12 provides free math and science curriculum to schools worldwide. When we started Flexi, the question wasn't whether to add AI tutoring. Every edtech company was doing that. The question was whether we could build one that a school district would actually let near its students.
That's a harder problem than it sounds. An AI tutor for K-12 has three users with competing needs:
1. Student who wants an answer right now
2. Teacher who needs to know the tool isn't undermining their classroom
3. District that has to answer for student data privacy and content safety.
Most competitors we analyzed optimized only for the first one.
What the research told us
The dominant emotion when students got stuck wasn't confusion. It was embarrassment. Students would rather stay stuck than ask a question that made them look behind.
This was later confirmed in the wild by an independent study (more below). A teacher described her strongest but painfully shy students:
"[They] by far benefited the most from having something that they could refer to without bringing attention to themselves. [They] came out of their shell more than I've ever seen."
Design implication: the tutor's first job is to be a judgment-free place to ask. Everything else builds on that.
That bet was later validated in an unexpected way. CK-12's conversation-pattern research found that many high-growth students began with casual, off-task chat ("Do you know that I'm good at Pacman") that evolved into real academic inquiry, and that students asking questions in their own voice, slang, misspellings, other languages, built the trust that made them willing to admit confusion. Off-task conversation isn't always a distraction. Sometimes it's the on-ramp. Flexi is designed to respond warmly to student voice and then steer back to learning, rather than shutting casual talk down.
Decision 1: An answer is the start of the interaction, not the end
If Flexi just answered questions, it would be a homework-completion machine, and teachers' single biggest fear about AI (overreliance, cited more than any other concern in later research) would be justified.
So the core loop is answer → check understanding → stretch further.
Explain the concept, immediately verify with a low-stakes question, then offer one step harder. Teachers later named this directly: "The 'Challenge Me' feature aided my students to go further with their current learning."
Did the scaffolding actually change behavior? Yes, and we can measure it. CK-12's analysis of Flexi conversations across the 2024–25 school year found that among students who asked Flexi for help during adaptive practice, direct answer-seeking questions dropped from 72% at their first query to 52% by their eighth, while deeper academic-learning questions rose from 11% to 26%. Students who progressed from "what's the answer?" to "why does this work?" also showed growth on practice scores. The scaffolds don't just deflect answer-grabbing. They teach a better way to ask.
Decision 2: Show the AI's uncertainty instead of hiding it
The instinct with a student-facing AI is to make it sound confident. We did the opposite: Flexi surfaces when it's uncertain and grounds responses in CK-12's curriculum library rather than open-ended generation.
A hallucinated answer delivered confidently to an adult is annoying. Delivered to a 12-year-old studying for a test, it's a genuine harm.
Teachers understood the stakes without prompting; one told researchers:
"I am concerned that some students can put too much trust in AI as offering error-free explanations... and not understand that AI can be wrong at times."
Decision 3: The teacher is the most powerful user in the system
Teacher trust was the adoption bottleneck, so we built controls rather than just reports: teacher-managed toggles like "Pause Flexi" and "Rephrase Response," and visibility into what students were asking (the "Asked on Flexi" surface).
The independent study confirmed the payoff. Teachers used Flexi as a "first responder" for routine questions so they could focus on students who needed more:
"During opportunities of review and extension, I can attend more to the questions and to the students who need more help. Flexi is like an additional tool or staff in the classroom."
And the transparency worked both directions: "With Flexi, both I and my students can clearly see progress and areas where they're struggling. This transparency keeps everyone more accountable."
Designing for Minors: The constraints that shaped everything
Content grounding: Responses draw on CK-12's curriculum library, which narrows what the model can get wrong and keeps answers aligned to what the class is actually studying. Teachers noted it "worked best when paired with existing curriculum resources like CK-12."
Privacy: Student interactions are subject to COPPA and FERPA.
Accessibility as a first-class need: The study found Flexi's text-to-speech and speech-to-text accommodations mattered most for special education students, and Spanish-language support made it "a powerful tool" in bilingual schools.
Age range: A K-12 tool spans a 7-year-old and a 17-year-old. We got this partly wrong; see below.
What an independent study found
LeanLab Education, a nonprofit that runs codesign research in K-12, conducted an independent seven-week, multi-method evaluation of Flexi: 10 middle and high school math and science teachers across five states, with pre/post surveys, user diaries, and focus groups.
What worked:
- Teachers reported increased student engagement, confidence, and curiosity, strongest among independent learners and shy students
- Teachers saw time savings and new insight into how students think through problems
- Teacher attitudes toward classroom AI shifted measurably positive over 7 weeks, and several said Flexi opened honest conversations with students about responsible AI use: "When they use it... they don't try to hide it from me, they're excited about it"
- Clear value for differentiated learning: special education students, struggling readers, English language learners
What didn't (and what we did about it):
- Students prompt like they search. Many treated Flexi like a search engine or just photographed problems, expecting instant answers instead of a conversation. \
- Design response: we built a more robust suggested follow-up foundation, and built prompt libraries to educate students on the best way to prompt LLMs and set them up to create their own
- Reading level was a barrier for struggling readers and multilingual learners.
- Design response: Flexi was built to auto-adjust to reading and learning levels over time. Adaptive practice scores, grade level, and more are factored in to how Flexi responds.
- The tone read young to older students. One teacher: it was "a little in your face and enthusiastic... an elementary school kid would love that." This confirmed what we'd seen with early avatar designs and drove the shift toward a more neutral voice and visual identity for high schoolers.
- Teachers' top concerns were overreliance on AI (mentioned 7×), loss of critical thinking (5×), and academic integrity (4×). These map directly to the design decisions above, and they're why the answer-first chatbot pattern is wrong for this domain.
The study also outlined a path to ESSA Tier III "Promising Evidence" qualification through a larger correlational follow-up, the evidence standard districts actually procure against.
What thousands of conversations taught us about measuring an AI tutor
CK-12's own research team analyzed conversation patterns at scale ("The Dialogue that Drives Learning with AI," 2025), pairing thousands of student conversations with adaptive-practice performance across a full school year. Two findings changed how I think about designing and measuring products like this:
The engagement paradox. High usage with shallow engagement predicted poor outcomes. Some declining students submitted hundreds of queries; they copy-pasted homework questions with zero follow-ups, or used Flexi almost entirely for non-academic chat as an avoidance mechanism. Usage metrics alone, sessions, minutes, message counts, are misleading indicators of learning. Question quality and follow-up behavior are the real signal. For a learning product, this is the difference between designing for engagement and designing for growth, and it's why "reward depth over efficiency" became the operating principle.
Conversation patterns are diagnostic. The research identified distinct behavioral profiles: productive strugglers, learned helplessness ("just give me the answer," refusing scaffolds), homework processors, even high performers gaming multiple AI tools at once. Each needs a different intervention, and the conversation itself is where you can detect them early, before the pattern hardens into a habit. That insight points at where AI tutoring goes next: surfacing these signals to teachers so they can intervene, which connects directly to the learning-analytics work in our Efficacy Studies (https://info.ck12.org/efficacy-studies).
