IA + UX Research | Trazur Courses

Researching AI and UX where connectivity can’t be assumed.

This project explores how AI can support UX research and redesign work on Trazur Cursos, without replacing human judgment or real-user validation.

1. Diagnosis
2. Comparative research: heuristics + AI
3. Findings & persona
4. Prototyping & redesign proposal
5. Validation with real users

This was my thesis project. I led the UX research, AI-assisted analysis, and design of the proposal, working alongside Renzo Morandi and under the guidance of Alejandra Capocasale — this case documents the full team’s work.

Scope: an academic project, validated through heuristics, AI, and prototyping — not yet with real users from the target audience. The percentages in this case are redesign targets, not measured results.

A functional platform, but with friction for part of its core audience.

Trazur Cursos had confusing navigation, dense copy, weak visual feedback, and real friction at signup and checkout. For a rural producer with low digital literacy and an unstable connection, any one of those could mean dropping the course altogether.

The direction was clear from the diagnosis: simplify the information architecture, write accessible microcopy, surface clear navigation states, and add a contextual assistant — using AI to speed up analysis, always under human review (more in Method and Key Insights).

Understanding how AI can improve the UX of a rural e-learning platform for livestock traceability.

Specific goals

• Explore applications of AI in UX research.
• Evaluate AI tools for usability analysis.
• Propose an evidence-based, workable redesign.

Early signals (diagnosis)

• High bounce rate on course pages.
• A high share of incomplete signups.
• Deep scrolling and disorientation, visible in Clarity’s heatmaps.
• Friction in the checkout flow and difficulty understanding copy, based on exploratory Synthetic Users simulations.

Redesign targets

Impact goals set before the redesign — not measured results (see Expected Impact).

A comparative strategy: traditional methods and AI tools, always read together.

We compared traditional UX research methods with AI-assisted tools to assess the platform. Each approach surfaced a different kind of evidence.

Nielsen heuristics. Identified consistency, navigation, and feedback issues — the foundation of the initial diagnosis.

Google Analytics. Showed a high bounce rate and erratic clicks on course pages.

Microsoft Clarity. Heatmaps revealed deep scrolling and disorientation, especially on mobile.

Synthetic Users. An exploratory simulation: it generated hypotheses about checkout friction and comprehension issues — useful for prioritizing, not a substitute for real behavioral analytics or real-user testing.

Benchmarking. Compared against similar platforms, Trazur showed structural gaps in navigation and usability.

Can AI speed up a heuristic evaluation without losing depth?

To answer that, I compared a manual heuristic evaluation against the same exercise run through three AI models — ChatGPT, Gemini, and LLaMA 3 — looking at what each one caught, what it missed, and how deep the analysis went.

The manual review stayed the most accurate and context-sensitive, grounded in direct experience. The AI models sped up the process and organized findings well against known heuristics, but showed real limits: ChatGPT tended to miss real user barriers; Gemini’s analysis was more superficial; LLaMA was technical but repetitive and less tailored. The honest takeaway: AI speeds up the first pass — it doesn’t replace the judgment of the person doing the research.

Combining traditional methods and AI tools surfaced clear patterns of confusion and critical friction.

• Confusing, unintuitive navigation.
• Dense copy and inaccessible microcopy.
• Missing visual feedback — errors, loading, states.
• Excessive scrolling on mobile, causing disorientation.
• Friction at signup and checkout, leading to drop-off.

Turning insights into an archetype to prioritize the redesign.

Confusing navigation, dense copy, weak feedback, signup and checkout friction: from those insights, we synthesized an archetype persona, an end-to-end journey, and an ideal flow for the critical path — sign up, take the course, get certified. These artifacts guide the redesign and its priorities.

Juan Pablo Techera

Family livestock producer.

• Age: 53
• Gender: male
• Education: completed lower secondary school
• Digital literacy: low
• Location: Tacuarembó, Uruguay
• Primary device: mid-range Android phone, intermittent connection

Biography

Juan Pablo is 53 and lives in Rincón del Sauce, Tacuarembó. He runs a family livestock operation with his wife and son. He has hands-on knowledge of the land but little formal tech training. He gets online through his Android phone, with an unreliable connection. He’s self-taught, and values tools that let him keep learning without stepping away from his daily work.

Goals

• Meet traceability requirements without relying on others.
• Learn to log and check livestock data independently.
• Build the skills to pass this knowledge on to his son.

Frustrations

• Platforms that are slow or fail to load on a weak signal.
• Long copy with no practical examples.
• Signup steps that require outside help.

Interests

• Efficient livestock production within regulation.
• Technology applied to agriculture.
• Accessible training for rural workers.

I wish someone would walk me through it — this kind of thing, I get lost easily.

Juan Pablo Techera, family livestock producer

Persona “Juan Pablo” — original work, thesis project, 2025.

Juan Pablo’s journey, from search to certification.

The map covers six stages, tracking Juan Pablo’s actions, thoughts, pain points, and opportunities at each one, from the first click to the downloaded certificate.

Most significant pain points

• During search: similar-sounding names and unclear search results.
• At signup: the form stalls registration, with no visible help in the moment.
• During the course: no clear progress indicator, and a dropped connection forces learners to repeat modules.
• At assessment: the certification process is confusing, with unclear steps to complete the course.

Every redesign decision responds to a concrete finding — not an aesthetic preference.

Three fidelity levels to test hypotheses before committing to a final visual direction.

1. Low fidelity (Balsamiq)
   Defined the core flows — signup, catalog, checkout — with no visual distractions.
2. Mid fidelity (Relume)
   Tested hierarchy, simplified navigation, and accessible microcopy.
3. High fidelity
   Brought together a visual design with a progress bar, clear completion states, and a contextual virtual assistant (chatbot) to resolve questions in real time.

trazur.uy — signup

trazur.uy — course progress

trazur.uy — confirmation

One unified flow, with short steps, visible feedback, and support at the critical points.

Each step in the flow responds to a specific finding from the sections above, and together they support the redesign targets.

• Homepage with direct access to courses and clear hierarchy.
• Course page with benefits, requirements, and a single CTA.
• Signup broken into short steps, with visible status.
• Clear checkout and purchase confirmation.
• Course progress bar with a contextual assistant for questions.
• Certificate accessible on completion.

Ideal flow — original work, thesis project, 2025.

A design hypothesis grounded in heuristics and prototyping — not yet validated with real users from the target audience (see Next Steps).

How this would be measured, not just what’s expected.

The quantitative targets are already defined in Goals & Expected KPIs. There’s no real data confirming them yet — the proposal hasn’t been deployed to production or tested with real users.

Measuring real impact means shipping the redesign and observing it in use: Google Analytics and Microsoft Clarity for bounce rate, signup time, and navigation behavior; course completion and certificate downloads as completed tasks; and usability testing with the target audience to validate comprehension and satisfaction. Without that cycle, these targets remain a design hypothesis, not a result.

What this project can’t claim yet.

• Synthetic Users is a hypothesis-generation tool, not a substitute for real behavioral analytics or testing with real people.
• Usability testing with real rural producers from the target audience hasn’t happened yet.
• The AI analysis showed limited contextual sensitivity: in some cases it missed real barriers or produced surface-level readings (see AI-Assisted Analysis).
• This is an academic project, not a production product — the redesign targets remain unmeasured (see Expected Impact).

Turning this redesign proposal into a validated, measured improvement.

The complete visual summary of the thesis project.

Heuristics, method, prototypes, and conclusions, in one poster.