Niko Zarzani
← Selected Work
NZ-01 · SIDE A · '24

Generative AI for Corporate Education

Role
UX/UI · Product Design · PM
Product
COSMO
Year
2024
Platform
Web app · Mobile app

Cosmo is an AI-powered learning companion that transforms a company’s knowledge into personalized, on-demand learning experiences by combining verified organizational knowledge with each learner’s goals, preferences and knowledge gaps.

Download the full case study (PDF) ↗
Mobile app

On-demand learning that adapts its format, depth and tone to the learner and the moment.

1 / 8
The mobile experience: on-demand answers rendered as maps, text, audio and video.
Desktop web app
1 / 5
Project overview

Reimagining corporate learning in the AI era

Cosmo is an AI-powered learning companion that transforms a company’s knowledge into personalized, on-demand learning experiences.

Instead of serving static courses or generic AI answers, the platform combines verified organizational knowledge with each learner’s goals, preferences and knowledge gaps. It can generate articles, concept maps, audio, video, quizzes and guided questions in the format that best fits the learning moment.

Its maieutic approach does not simply deliver information: it uses dialogue and questions to encourage reflection, curiosity and lasting understanding.

The challenge

Moving beyond content delivery

Most corporate learning platforms are built around catalogs, courses and completion rates. Employees are often overwhelmed by choice but underwhelmed by relevance, while L&D teams struggle to personalize learning without losing control over quality, privacy and company culture.

The challenge was to create an experience that could adapt to the individual while remaining deeply grounded in the organization’s language, values and verified knowledge.

How might we design a learning experience that adapts to each user’s preferences and pace, while staying aligned with the company’s identity?
My role

Leading the experience from research to beta

As Senior Product Designer, I led the product experience from early research to beta validation.

I framed the problem space through interviews and field observation, defined the UX and AI interaction principles, designed the learner-facing web and mobile experiences and worked closely with developers on the underlying AI architecture.

I also contributed to MVP prioritization, privacy decisions, system sustainability and stakeholder alignment, while mentoring a junior designer and maintaining the overall design direction.

Product strategyUser researchService designInformation architectureAI interaction designUX/UI designPrototypingDesign systemTestingProduct management
Research

Understanding learning beyond content

Research combined contextual interviews with new hires, learners, HR managers and L&D leads with direct observation inside real training sessions across pilot organizations.

Four behavioral insights shaped the product.

  • Personal motivation matters

    Learners wanted content connected to their personal motivations, not only their job title.

  • Questions create deeper learning

    Learning through questions and reflection resonated more than traditional top-down instruction.

  • More content does not mean more relevance

    Employees were overwhelmed by the quantity of available content but rarely found it relevant.

  • Personalization requires governance

    HR teams wanted personalization while retaining control over organizational knowledge and values.

1 / 4
Personas distilled from interviews and field observation across pilot organizations.
Design principles

Crafting the learning DNA

The research was translated into four principles that guided both the product experience and the AI system.

  • Personalization with editorial soul

    Use AI to scale individual learning without reducing the experience to generic automated content.

  • Multimodal by design

    Let learners move between text, maps, audio, video and interactive activities depending on their context and preferences.

  • Company identity over generic knowledge

    Ground every answer in verified sources that reflect the organization’s language, culture and expertise.

  • Progressive, goal-driven interaction

    Reduce cognitive load through gradual onboarding, proactive guidance and clear learning objectives.

The experience

From a static LMS to a proactive learning companion

Cosmo was designed around curiosity rather than course catalogs.

A progressive onboarding experience begins to understand the learner without requiring a long initial questionnaire. A personalized discovery feed surfaces relevant topics and knowledge gaps, while an AI mentor helps users explore questions through conversation.

Daily learning loops turn broad development goals into small, achievable actions. Progress indicators and lightweight gamification make growth visible without distracting from the learning itself.

The result is an experience that actively guides the user instead of waiting for them to navigate a conventional LMS.

"The platform doesn't give neutral answers. It starts from a knowledge base that reflects the company's identity."
AI architecture

Where AI meets organizational identity

The experience was supported by a modular three-layer architecture.

Four primary agent functions coordinated by a central orchestration layer over company and learner knowledge.

Three knowledge layers (Company Knowledge Base, Learner Knowledge Graph and Generative Output Layer) feed a central orchestration layer that coordinates four specialized agents: Plan, Create, Mentor and Evaluate.

Specialized AI agents handled different responsibilities: planning learning paths, generating content, mentoring learners and evaluating assignments. An orchestration layer maintained context across the experience, coordinated the agents and applied shared guardrails for tone, privacy and compliance.

Trust and privacy

Designing personalization without surveillance

AI-generated learning content needed to feel relevant, but also reliable and safe.

The system was grounded in verified internal learning assets rather than unrestricted internet knowledge. Sensitive analytics were protected through role-based access, while identifiable personal data was processed within a private environment and kept inaccessible to external parties.

This allowed personalization to support the learner without turning the experience into opaque employee surveillance.

Trust was not treated as a compliance layer added at the end, but as a core part of the product experience.
MVP and prototyping

Turning the vision into a testable product

Before moving into detailed design, features were prioritized through co-design workshops and a MoSCoW framework across three user groups: learners, L&D and HR managers and knowledge-base curators.

The first prototypes explored progressive onboarding and learner profiling, personalized knowledge discovery, conversational AI interactions, learning goals and progress, multimedia content generation and feedback and evaluation.

Low-fidelity flows were progressively translated into interactive web and mobile prototypes and validated with users and technical stakeholders.

1 / 3
From early flows to interactive prototypes: progressive onboarding, discovery and conversational mentoring.
Beta validation

Testing the experience in the real world

The beta was presented at the We Make Future innovation fair, where visitors could interact with the product in an open, hands-on environment. This provided early behavioral observations, qualitative feedback and a first comparison with more traditional, static learning experiences.

Figures below are beta and pilot results from early deployments, not universal or independently audited outcomes.

  • +32%
    Average daily learning interactions
  • +47%
    Content completion compared with the legacy LMS experience
  • −65%
    Average time required to answer training-related questions
  • 86%
    Positive feedback on the tone and style of AI-generated content
Hands-on beta testing at We Make Future: spontaneous interaction, qualitative feedback and early behavioral signals.
Product iteration

Learning from behavior, not assumptions

From generic to context-aware answers

Problem
The first AI mentor responses were correct but often perceived as too generic.
Evidence
The initial average usefulness score was 3.6 out of 5.
Insight
Users expected vocabulary, examples and recommendations that reflected their organization’s specific way of working.
Intervention
Prompts were enriched with additional company context and feedback tags were introduced for useful, partially useful and irrelevant answers.
Outcome
After two tuning cycles, the average rating increased to 4.3 out of 5, with users describing the responses as more relevant to their learning context.

Reducing onboarding friction

Problem
The original onboarding asked ten questions before demonstrating the platform’s value.
Evidence
Twenty-two percent of users abandoned the process before completing their learning profile.
Intervention
The initial questionnaire was reduced to three optional questions. A conversational development-plan agent was introduced and additional information was collected progressively over time.
Outcome
The changes produced a 56% reduction in onboarding drop-offs and stronger early engagement.
Reflection

Designing an environment for lasting growth

The project’s success did not come from generating more content. It came from creating a learning environment that could feel intelligent, familiar and connected to both the learner and the organization.

Combining design thinking with AI system design allowed us to build a product that adapts over time without losing coherence, identity or trust.

Great AI experiences are not only about usability. They are about trust, curiosity and connection.
The team
The COSMO team and I at the product presentation, June 2024 🤖