AI UX · Product Design · Conversational Systems · 2026

Designing a conversational portfolio assistant

A contextual way into the work: visitors ask their own questions, receive evidence-bounded answers, and discover the most useful project to explore.

Role

Product Design · AI UX

Focus

Conversation · Retrieval · Governance

Artifact

Working product in this portfolio

Status

Live · continuously iterated

Design lens

A conversational access layer for the work

A small conversation can open a meaningful route through an otherwise linear portfolio. The design challenge was to interpret questions, preserve evidence boundaries, protect private work, and choose when to answer, link, or stop.

01Grounded answer
02Relevant case study
03Useful next path
04Human contact

Illustrative interface · reconstructed QA example, not a live transcript

Portfolio Assistant

Show me representative work.

An illustrative answer

Start with Par for Each for research-led enterprise product design, Citizen Portal for end-to-end service design, and DesignOps for organizational systems thinking.

Visitor question → Output: a concise, grounded answer based on portfolio evidence

The interface states below are reconstructions; the assistant at the end is the live product.

Core design work

What I actually designed

The interface is small, but the design surface is not. I designed five connected systems — from the visible experience to the information architecture and ongoing evaluation behind it.

↳ Select a design area to inspect the decisions and system behind it.

Experience design

The visible journey from first entry to evidence and human contact.

  • 01Quick actions reduce cold-start friction without turning the chat into a menu.
  • 02Answers stay intentionally compact so the chat orients rather than replaces the portfolio.
  • 03Links move the visitor back into evidence; escalation appears only after meaningful engagement.

01 · Problem

A portfolio has a fixed structure that often doesn't directly answer a visitor's individual question.

Case studies tell structured stories, but visitors may want to compare capabilities, trace a design decision, or assess role fit across several projects. I treated this mismatch as a design hypothesis to test with a working assistant — not as an established finding about every recruiter.

Design challenge

How might conversation provide individual access to the portfolio without becoming another general-purpose chatbot?

02 · Design

Conversation is navigation, not the destination.

I treated the assistant as a product surface with explicit interaction principles, not as a chat window attached to the site.

↳ Select a principle to see why it matters and how I translated it into the product.

Why it matters

A convincing answer is useless if it cannot be traced back to real work.

How I implemented it

Responses are assembled from retrieved portfolio evidence and are instructed not to invent projects, impact, qualifications, or responsibilities.

Exploring the interaction model

The first question was where conversation should live.

I explored four ways to give visitors conversational access to the work. The deciding factor was whether the assistant could stay useful without becoming a separate destination.

Exploring the interaction model

Conventional chatbot

Familiar, self-contained

01 / 04
Portfolio Assistant

Hi! I can help you explore Andreas’s work.

Tell me about your background.Which projects should I start with?How can I get in touch?
Ask a question...

Familiar, self-contained

Familiar interaction model. Too detached from the portfolio itself.

Compare all four directions
Exploring the interaction model
DirectionStrongest qualityDecision
Conventional chatbotFamiliar interaction modelToo detached from the portfolio itself.
AI command / searchFast information retrievalEfficient, but too search-like for a conversational experience.
Embedded conversationMaximum local contextStrong within a page; fragmented for questions across projects.
Contextual companionChosen directionContext and cross-portfolio explorationKeeps the portfolio primary while adding individual conversational access.

The contextual companion offered the best balance: understand where a visitor is, answer across the portfolio, and send attention back into the work.

The chosen interface

The final interface is deliberately quieter than the system behind it.

Familiar conversational mechanics, with context, navigation and escalation appearing only when they help the visitor.

Portfolio Assistant
Context · Par for Each
What did Andreas learn here?

Portfolio answer

Three rounds of usability testing shifted the design toward a Condition Modeler. Separating the end condition from iteration logic made the workflow easier to understand; clearer threshold wording helped people express when parallel execution should stop.

Read the Par for Each case study

Ask a question...

  1. Header and dismissal

    The name identifies the assistant; the close control returns attention to the portfolio.

  2. Page context

    A quiet line under the header identifies the current case study or reflection.

  3. Editorial answer

    A small label and flat text replace another large chatbot bubble.

  4. Inline navigation

    Validated portfolio links sit inside the answer, without a separate source layer.

  1. User message

    A dark, right-aligned message makes the visitor’s input distinct.

  2. Follow-up restraint

    At most one optional continuation. A distinct follow-up can accompany a link when it adds a useful direction.

  3. Message input

    A quiet input row and send control keep the next question within reach. At the conversation limit, input closes.

Core interaction model

·

More context underneath. Less machinery on top.

Visible context

Current page

Par for Each

"What did he learn here?"

Visitor sees: a normal question in a quiet chat interface.

Context priority

  1. 01

    Explicit question

    Project evidence

  2. 02

    Conversation anchor

    What Andreas did, tested, shipped.

  3. 03

    Current page

    Page-specific evidence

  4. 04

    Global profile

    What Andreas argues, explores, questions.

Grounded response

Usability testing showed that the interaction needed to separate the end condition from the iteration logic… Read the case study

No evidence?

The assistant says the portfolio does not establish it — and stops there.

Key-state storyboard

Seven states. One deliberately bounded experience.

From a quiet invitation and context recognition to evidence, uncertainty and exit: the experience is a sequence of designed states.

Key-state storyboard

Entry

A quiet invitation

01 / 07
01

The portfolio remains the primary experience. Opening a conversation is optional.

ANDREAS JOHANN

Product Design · UX Research

Par for Each

Designing an End Condition for Parallel Subprocesses

Making parallel execution easier to understand and control.

Ask about this project

The interaction succeeds when the visitor leaves with a clearer path into the work — not when the conversation lasts longer.

03 · Build

The product is small. The representation system behind it is not.

I translated the interaction model into a working system: intent routing, knowledge retrieval, behavioral rules, conversation state, and controlled navigation — all explicit enough to govern.

01

Question

Intent · locale · current page · conversation anchor

02

Route

Named domain or broad profile intent

03

Retrieve

Relevant portfolio evidence only

04

Frame

Scope · claims · language · link policy

05

Generate

Constrained answer from supplied context

06

Experience

Answer · project · follow-up · contact

Cross-cutting controls

Conversation stateVerified linksDomain separationPortfolio-only scope

The AI contract

01Knowledge boundary

Use portfolio evidence and preserve its evidence type. Do not turn reflections into project proof or invent work, impact, qualifications, or experience.

02Domain boundary

Do not merge adjacent workstreams simply because they share words such as scale, systems, or cross-team.

03Management boundary

Program leadership and DesignOps work do not become formal people-management claims.

04Navigation boundary

Only validated portfolio destinations are rendered as clickable links.

05Access boundary

Public visitors only receive public evidence; after unlocking, the assistant may use authorized detail. Links retain the project's password gate.

One product, multilingual behavior

Language is treated as part of the product behavior rather than post-processing. The assistant is designed to respond consistently within the portfolio's localized experience.

↳ Select a language to inspect the same product behavior in a different localized expression.

What kind of product designer is Andreas?
A product and UX designer who works from research and interaction design through delivery, with additional depth in AI-native UX and DesignOps.

Public access, deliberately scoped

Without unlocking, the assistant uses public evidence and project summaries. It must not disclose protected research, metrics, or implementation details; deep links still lead through the normal project access gate.

04 · Evaluate & Govern

AI quality is not a one-time usability check.

Changes to the model, prompt, retrieval logic, knowledge base, or portfolio itself can all alter behavior. I therefore treated evaluation as an ongoing governance loop rather than a final validation step.

Retrieval

Did the system bring in the most relevant evidence?

Representation

Did the answer portray the work through the right professional frame?

Behavior

Did it stay in scope, use the right depth, language, and next action?

From QA failure to product principle

Retrieval is not curation. A correct answer that deflects is still a failure. Example: "Show me representative work."

Before

"Show me representative work."
You can explore Andreas's representative work by visiting his portfolio.
Source · Work indexFollow-up · Show me representative work

What failed

Deflects to navigation instead of curating. Provenance row adds noise. Follow-up prompt is redundant.

After

"Show me representative work."
Start with Par for Each for research-led enterprise product design, Citizen Portal for end-to-end service design, and DesignOps for organizational systems thinking.

What changed

No visible provenance row. No redundant follow-up. Direct paths into evidence — the assistant chose what matters.

Privacy-conscious telemetry

routeevidence pathlink clickno-evidencecontactlatencylimit reached

No raw prompts or answers stored

Systematic test set

↳ Select a test domain to inspect a representative question and the expected behavior.

"Would Andreas be a good fit for a Senior Product Designer role?"

Expected behavior

Prioritize product-design craft and representative project evidence.

Continuous evaluation & governance

01

Observe

02

Classify failure

03

Trace source

04

Adjust

05

Regression test

06

Monitor

Hallucination

Factual claim not supported by portfolio evidence

Retrieval failure

The wrong evidence enters the answer

Representation drift

Facts are defensible, but the professional framing is wrong

Domain bleed

Evidence from adjacent workstreams gets merged

Behavior regression

Language, answer depth, link behavior, or persistence changes unexpectedly

Still to validate: Routing checks and manual QA do not yet establish improvements in visitor comprehension, hiring decisions or contact conversion.

05 · Iterate

The assistant evolved as the portfolio evolved.

Each iteration changed less about the visual chat UI than about what the system understood, retrieved, and was allowed to infer.

Signal

The assistant needed useful contextual entry points even before a visitor unlocked protected projects.

System change

Added access-aware evidence, page-specific starter questions, wider case curation, and a soft transition to contact before the ten-turn limit.

Not every important iteration changes the visible UI. Retrieval, evidence boundaries, and permissions are often invisible.

Reconstructed interface

Portfolio Assistant
Context · Par for Each
What did Andreas learn here?

Portfolio answer

Three rounds of usability testing shifted the design toward a Condition Modeler. Separating the end condition from iteration logic made the workflow easier to understand; clearer threshold wording helped people express when parallel execution should stop.

Read the Par for Each case study

Ask a question...

Reconstructed from the product decisions and behavior of each iteration.

When the portfolio changes, the assistant's information architecture has to change with it.

06 · Lessons Learned

The most valuable findings came from failures that looked plausible.

01

Correct information can still create the wrong representation.

An early answer to “How does Andreas conduct UX research?” drifted toward human-in-the-loop systems and AI trust. The content was defensible, but it framed the profile through the wrong lens. The model was not the core problem; the information architecture was weighting the wrong things.

02

The assistant became a mirror of the portfolio itself.

Each iteration changed less about the chat UI than about what the system understood the portfolio to mean. The assistant exposed which signals in the portfolio carried the most representational weight.

03

Semantic proximity is not conceptual equivalence.

UX Consistency and DesignOps both contain signals like scale, systems, cross-team, and quality. Live testing showed that shared vocabulary is not enough reason to merge evidence. Adjacent domains need explicit boundaries.

Try the live product

Don't take the case study's word for it.

Ask the assistant something specific about my work and see how the current system retrieves evidence, frames the answer, and navigates you deeper into the portfolio.

Next Project

Scaling Approval Workflows

Increasing process efficiency while ensuring compliance for SAP Build Process Automation

Read case study →