Bringing on-the-job learning into education
How do students learn on the job inside a micro-role apprenticeship program? A qualitative study with 12 interviews, and what it means for platform design.
Select a stage to see how learning unfolded.
The problemReal work is how students learn, but classrooms rarely offer it
Universities increasingly use project-based learning, but a gap remains between bounded academic exercises and the demands of professional practice. The gap matters more now because AI is taking over many entry-level tasks, such as simple coding and documentation, that used to be the "onboarding ground" for novices. New graduates are expected to arrive with higher-level know-how that once took years of junior-level work to build.
The challengeRealistic projects are unstructured, and structured ones aren’t real
The challenge: authentic settings like open-source projects and internships are often unstructured. Tasks are unpredictable, and nothing guarantees a suitable task for every learner and skill level. Education needs the opposite: predictable, scaffolded pathways that work for every student. Realism and structure rarely come together.
The opportunity: the micro-role apprenticeship model breaks a large project into a hierarchy of well-defined micro-roles, giving learners a structured pathway while keeping their work part of a larger real project. An earlier proof of concept had only covered static web pages, and no one had studied how learners actually experience learning this way, or tried it across full-stack development.
The approachInterviewing 12 people inside a micro-role program
The program
We built an in-person, community-oriented skill-building program in a research lab, developed iteratively since 2020. It has three micro-roles. Components is the entry level, where learners build reusable UI pieces and learn HTML, CSS, TypeScript, and Angular. Containers coordinates data between the backend and the frontend. Applications is the lead role, where members mentor others and architect the application. Learners master one micro-role per quarter, work in pairs, and finish each two-week cycle with peer-reviewed GitHub pull requests and lead approval.
Why interviews, and why this analysis
- The question was about experience. No one had examined what learning on the job feels like inside a micro-role program, so interviews let participants describe it in their own words.
- The goal was design, not just description. We wanted to find the cognitive and social friction points a digital platform would need to address, and those only show up in how people talk about getting stuck.
- Semi-structured kept it comparable. Every hour-long interview covered the same topics (tasks in each micro-role, problem-solving, resources, and mentoring for leads) with room to follow what participants raised.
- Inductive thematic analysis let the stages emerge. We coded the de-identified transcripts without fitting them to a model chosen in advance.
Who we interviewed and how we analyzed
We interviewed 12 participants from 2024 to 2025. Eight of them had served as leads for at least a quarter, so we heard from both learners and mentors. Authors independently open-coded the transcripts, then clustered the codes into higher-order categories through iterative discussion.
My role
As Research Lead, I:
- Developed the interview protocol and materials, and ran a subset of the participant interviews
- Coded and analyzed the qualitative data using structured coding frameworks and thematic analysis
- Synthesized findings into recommendations for the program and the platform design
- Led quarterly research planning and coordinated the team's weekly priorities
The resultsFive stages, from lost to independent
Participants moved through five stages as they learned on the job. Hover over or tap a stage to read more.
Disorientation
Many learners started out completely lost.
Tasks began right away, so learners scanned each micro-role's supporting resources to build a mental map of the task and how it fit the larger system.
Imitation
First tasks followed worked examples.
Learners valued tasks broken into small subgoals and followed step-by-step walkthroughs, sometimes copying example code without fully understanding it.
Learning through differences
Learners compared new tasks with earlier ones.
As they gained familiarity, they outgrew the walkthroughs and looked for what was similar and what was different between tasks, then adapted earlier work.
Help-seeking
When stuck, they looked for targeted help.
They searched documentation and Stack Overflow, used AI tools, and asked peers and leads about specific problems or error messages.
Mentorship
Check-ins helped hesitant learners recover.
Some learners felt bad about "pestering" their leads and got stuck. Leads who checked in on their team helped them get moving again.
What worked
- Micro-roles gave a manageable, motivating pathway, in contrast to earlier self-teaching that felt like falling into a "rabbit hole."
- The hierarchy gave a tangible sense of growth from one role to the next.
- Working in pairs and teams supported peer learning.
- Leads were in the same shoes as learners, and the bounded roles let them give targeted feedback.
- The work felt real because it could be merged into larger projects.
What didn't
- An overemphasis on task completion could hinder deep learning. Learners sometimes prioritized finishing over understanding, and leads questioned whether they grasped the concepts.
- With productivity as the main measure, learners were tempted to lean on AI tools for quick completion.
- Community structure could make motivation fragile, since disengaged leads or peer comparison affected how learners felt.
From findings to designA platform built to counter shallow learning
The findings directly shaped the requirements for a digital platform, a "Khan Academy for experiential on-the-job learning" built around micro-roles, subgoals, and walkthroughs.
- Counter shallow understanding. Each walkthrough blends planning steps, concept steps, task steps, a review step with a code diff view, and AI-generated quizzes, so learners reflect before and after the task, not just finish it.
- Support learning through differences. The dashboard shades a step differently when a learner has already completed a step of the same type, so new kinds of steps stand out. A planned AI agent would flag what is new and what is familiar in a task.
- Recreate near-peer support. Planned features include small cohorts working on the same micro-roles and code review among peers, near-peers, and AI.
Impact
The study gave the field the first qualitative account of how learners experience micro-role based learning in a real-world setting, and it fed directly into the platform's design. It was published at CHI 2026 (Extended Abstracts) and funded by the National Science Foundation.
“… like you’re genuinely in a mini company.”
ReflectionWhat learning in tiers taught me
Working on this study showed me how much firsthand experience matters, and how well it works when it comes through tiers of mentorship. In the findings, near-peer mentors helped students get unstuck. I saw the same thing in my own role: I learned the work by doing it alongside people at different levels of experience. It also made me careful about how learning gets measured, since the study itself showed that counting finished tasks can hide whether someone understands the concepts behind them.
Read the paper Back to portfolioFeel free to reach out, I promise I'm friendly!
sssalunk@ucsc.edu