ALEPH ALPHA
Feb 2026 - NOW
Designing ELAI, an agentic learning system for complex AI workflow tools
A support agent inside Aleph Alpha's AI products that spots where users get stuck and guides them to the features they skip, without leaving the task.
ROLE
Product Owner & Designer
User Researcher
COLLABORATORS
Supervised by
Aleph Alpha Learning Solution Team
IMPACT
Started as my master’s thesis, pitched at Aleph Alpha’s HQ, now in MVP development with their learning team.
BUILT WITH
Claude Code
Vercel AI SDK
Paper
Storybook
CONTRIBUTIONS
Research
Benchmarked adoption platforms against enterprise AI needs.
Product Strategy
Shaped ELAI to fit a sovereign AI provider's go-to-market.
Interaction Design
Designed how the agent nudges, explains and points the way.
AI Harness
Set up the prompts, context, tools and model behind it.
I. CHALLENGE
An AI tool engineers rely on,
but barely explore.
Requirements engineering can take up to 37,500 hours per development project at an automotive supplier like Continental. Aleph Alpha’s Semantha speeds this up by matching each new customer specification against requirements from past projects.
For most engineers, semantic matching & reviewing is a new way of working with AI. Interviews with Semantha’s UX designer and product manager showed that the product is complex enough to need real support, and that support isn’t built in.
User Workflow
Upload requirements
User
Compare against library
AI
Review matches
User
Update library
User
Screen - New analysis in Semantha
Advanced features get skipped
“They want to save time.”
UX Designer, Semantha
Unexplained AI concepts
“Users don’t really understand the terminology
behind some features.”
UX Designer, Semantha
Missing support
“It's not something you can just throw over the fence to them.”
Product Manager, Semantha
Help center
Bring learning into the flow of work.
With the Academy, an LMS running on Docebo, Aleph Alpha already has a learning solution to educate its users and customers. But at a kickoff in February 2026, the Learning Experience Solutions team told me that leaving the product to find help in the Academy means too much context switching.
They wanted shorter learning units that live inside the products. My task was to explore new approaches, research the state of the art and turn it into a concept with a working prototype.
3 MIN
AI Governance & Compliance 101
Uploading Documents to AI Chat
3 min
EN
4.8
The ask
A second entry point for learning inside Aleph Alpha's products, starting with Semantha.
Constraints
Run on Aleph Alpha's own model, Kolibri, and infrastructure, including on premise. Use their new design system.
Success looks like
A working prototype that lets Aleph Alpha's teams decide whether to build it.
How might we…
… help engineers learn Semantha's deeper features without leaving their task?
II. SOLUTION
So I built
ELAI, a support agent
that nudges users towards the features they skip.
Screen - New analysis in Semantha
Marks user struggles
Highlights a feature the user hasn’t tried yet and asks if they want to know more.
Explains on demand
Shows how the feature works right in the task, with its own cursor and a short explanation
Links to Academy
Nudges users to go deeper with an Academy lesson that matches their task.
But how does the agent
know when to step in, and when to stay quiet?
So I designed a reasoning chain the agent runs before it speaks. It compares what someone actually does with the ideal task flows we mapped for the product, and weighs why the two differ. A nudge only appears once the reason to stay quiet has been ruled out.
Checked usage of
Similarity threshold
kept at the default 50% in all 20 analyses
↳
Tooltip
never opened
Ranked three explanations
Why would someone never touch it?
① Never noticed it
High
no interaction at all, tooltip never opened
② Unsure what it does
Low
would show as tooltip opens or slider resets, none seen
③ The default fits
Medium
plausible, most users keep it, so test it
Tested ③ against past
Matches
↳ 9 of 40 auto-accepted matches corrected later
③ The default fits
Ruled out
Highligh on
Similarity threshold
shows the next time the user opens a new analysis
How reliable are the answers?
Aleph Alpha trains its own model, Kolibri 1, to decline when there is no source. ELAI is built to run on Kolibri and holds the same bar: every answer is grounded in Aleph Alpha's own knowledge. It retrieves the matching passages from tutorials, documentation, the glossary and Academy courses, answers from them and cites the source. If nothing fits, it says so, offers the closest lesson and logs the question as a gap for the product team.
Users open it from the corner or with ⌘K, without leaving the task. ELAI already knows the screen and the step they're on, points to the right field and links to the matching Academy lesson when they want to go deeper.
KNOWLEDGE
Product videos
Documentation
Glossary
Academy courses
CONTEXT
On-screen
Task progress
User profile
But how do we adjust, test and evaluate the agent?
So I built an editable harness around ELAI. Its sources, skills and rules change without touching code, and every change runs against real questions. Each answer shows the sources and rules behind it, so we see what works before users do.
Test & evaluate
Knowledge
Skills & rules
Coming soon - under development
III. RESEARCH
How are other support & learning agents built?
Before designing, I took apart support and learning assistants that live inside a product, to see which patterns already work. I looked at what triggers them, what they know about the user and the screen, and how far they go on their behalf. The map sorts them by two of these questions, who starts the help and who ends up doing the task.
PRODUCT Positioning
The top right was nearly empty, so that is where ELAI sits. Engineers rarely ask about features they don't know exist, so ELAI steps in on its own. It teaches while it helps, so the skill stays with the user, which fits a company that sells sovereign AI.
↑ BUILDS THE USER'S SKILL
↓ DOES THE TASK FOR YOU
← WAITS TO BE ASKED
STEPS IN ON ITS OWN →
Brilliant
Hint beside the task when people get stuck.
Amplitude AI Assistant
Spot struggle in usage data.
WalkMe
Offer help before people ask.
Foldspace
Read the screen the user is on.
Intercom Fin
Test answers before users see them.
Microsoft Learning Agent
Tailor learning to the user's role.
ELAI
Steps in on its own and teaches.
PERSONAS
Two users, one agent.
Most support agents need two groups of users. End users get the help, and a team behind them sets it up, like the builders in WalkMe or the support teams who configure Fin. At Aleph Alpha, that team is their AI solutions engineers, who bring each solution to a customer and hand it over. So I designed personas to shape ELAI's user experience around the end user and its harness around the solutions engineer.
End user
Operator
WHO
Requirements engineer at the customer, new to working with AI
AI solutions engineer at Aleph Alpha
JOB
Match customer specs against past requirements in Semantha
Set up a solution at a customer and hand it over within weeks
GETS IN THE WAY
Skips features they don't know exist
More setup time, a tutor that says something wrong
NEEDS
Help at the step, without leaving the task
Fast setup, answers they can check before users see them
MEETS ELAI
In the product
In the harness
III. DECISIONS
How I designed an agent that builds
competence, not only resolving.
Decision 1
Skill profiling vs. task tracing
Teaching only works where something is missing. So before ELAI can teach anything, it has to find a user's skill gaps.
SKILL PROFILING
My first approach tried to score each user's AI competence with the product and build a profile from it. The estimates weren't reliable enough to act on, which matches what research finds for LLMs as learner models (Hooshyar et al., 2025).
User 0417
Regular
SKILLS
LEVEL
Prepare requirement data
Configure an analysis
Judge a match
Correct the results
BEFORE
TASK TRACING
So I mapped the product's task flows, which show ELAI the expert path through a task. Now it compares a session against that path and steps in where the gap costs the user something, like matches they later have to fix.
Task flow
Steps
Friction
⌄ Start an analysis
5
⚠ 2 steps
1
Enter a
Name
◆ Suggested
2
Upload the
Excel file
◆ Suggested
3
Set the
Degree of similarity
Kept at the default in 20 of 20 analyses · ELAI steps in
⚠ 3
4
Narrow by
Tags
(optional)
Not used in 20 analyses · ELAI steps in
⚠ 1
5
Continue with
Next step
◆ Suggested
AFTER
Decision 2
Everything is a plugin.
To learn what makes ELAI work, I needed to change one part at a time and see the effect. So I built the harness from plugins, following the DeepSeek Harness: model, tools, skills, rules, knowledge and the learning framework each sit in their own slot. Swapping one, like Mistral for Kolibri, runs against the same test questions, so the results stay comparable.
ELAI
Kolibri-1
Kolibri-1
Aleph Alpha · sovereign · DE/EN
Mistral Small
Mistral · EU · open weights
Nemotron Super
NVIDIA · open weights · on-prem
Product
Semantha
3 screens connected
Tools
cursor_action
post_to_slack
Skills
/walkthrough
/discover-feature
/explain-term
Knowledge
Knowledge
Videos
Product documentation
product-documentation.md
Vocabulary
Teaching
Decision 3
Engineer-maintained vs. customer-owned
In most support tools, one team sets up the agent and keeps running it. At Aleph Alpha, solutions engineers hand over the knowledge once a solution is integrated, then move on to the next customer, and the domain experts who stay don't code. So I split the work into two phases, each built for the person who does it.
INITIAL SETUP
ELAI rolls out together with the solution. The solutions engineer connects the product's sources, and ELAI drafts the task flows from documentation, research and tracking, with a source for every step. The engineer confirms or corrects each step. Before hand-over, ELAI answers a set of real questions from the proof of concept, and the tutor only goes live once it passes.
ENGINEER
Connects product and sources
ELAI
Drafts task flows with sources
ENGINEER
Confirms or corrects each step
ELAI
Runs the test set from the pilot
ENGINEER
Hands over to the customer
ELAI
Sends the weekly digest
OWNER
Reviews the proposed fix
ELAI
Tests it on past questions
OWNER
Releases the fix
WEEKLY REVIEW
After hand-over, a domain owner at the customer looks after ELAI. Nobody has to check it daily. Once a week, ELAI bundles the questions it couldn't answer and the steps that went stale after a release, each with a proposed fix. The owner reviews the fix, tests it against past questions and releases it.
IV. IMPACT
From concept to MVP development:
ELAI is entering its next phase.
Pitching ELAI at Aleph Alpha.
I had the chance to pitch the concept at Aleph Alpha’s headquarters in Heidelberg, to product managers, designers and the go-to-market team. The response was positive, and the discussion quickly turned to where ELAI could fit, from go-to-market to product development.
After the pitch, I teamed up with a member of their Learning Solutions team to take ELAI beyond my master’s thesis. The PRD, the GitHub repo and a deployed prototype are shared inside the company and updated as the MVP grows. Semantha was the first product ELAI learned, and we’re now talking to individual product teams to find the next ones.
Research for their team
ELAI started as my master's thesis. Its market analysis, user flows and personas went to Aleph Alpha's learning team as a starting point.
Pitched at their HQ
I presented ELAI in Heidelberg to product managers, designers and the go-to-market team. The talk quickly turned to where it could fit.
Now in MVP development
Together with their Learning Solutions team, we're building an MVP to test with real users.
Hi, I'm Joshua!
I design B2B products & features that grow with research that clarifies.
Best viewed on desktop
The case studies are built for a large screen. The mobile version is on its way.
Copy link for my desktop