Sascha Thiele for DuckDuckGo

01 · Creative brief

Search results, with privacy you can see.

The results page is where most people meet DuckDuckGo, and it's the one screen where the privacy promise is mostly invisible. My brief: make protection visible on the results page, make answers more trustworthy, and help people find the power features, all without making the page busier.

The insight

DuckDuckGo protects people after they click. The results page is the only place it can tell them before they do.

Tracker blocking, cookie pop-up handling and Duck Player all happen after you click, in the browser. On the results page every link looks the same, so the protection people trust DuckDuckGo for doesn't show up where they're choosing what to open.

What I noticed on today's page

Every link looks the same. A page with 0 trackers and a page with 25 look identical.
Search Assist reads as one block. Sources sit at the end, and when they disagree the answer doesn't say so.
!Bangs are hard to find. One of the best features is only there for people who already know about it.

Goals

Show what a click costs, in privacy terms, before the click.
Answers you can check: inline citations and an honest state when the sources don't hold up.
Teach !bangs at the moment they'd save someone a step.

Non-goals and constraints

No new data about people. Every signal describes a website, never a user.
Ranking doesn't change by default. Filtering by trackers is opt-in and visible.
One line per result at most. Details only appear when someone asks for them.

Success metrics

"What does the shield mean?"Correct explanation in moderated tests (n=8)≥ 80%

Citation useAnswers where a citation gets opened↑ vs. control

!Bang adoptionSearches that use a suggested bang↑ 2×

Browser installsFrom the "blocked in DuckDuckGo" prompttracked

Measured only with anonymous, aggregate counters: no user IDs and no linking of sessions, in line with how DuckDuckGo already measures.

Built on what DuckDuckGo already hasNo new infrastructure, just bringing existing protection forward to the results page.

Tracker RadarOpen crawl data behind every privacy line

Cookie Pop-up ProtectionShown on the result before you click

Duck PlayerVideo results say how they'll play

Search Assist & Duck.aiCitations, honest states, private follow-ups

Today (schematic)

1Results all look the same, so the privacy differences between sites stay hidden.

2The answer and its sources are separate. You can't tell which source backs which sentence.

3Bangs, Duck Player and Duck.ai are only found by people who look for them.

Proposed

1A one-line privacy summary on every result, with the full breakdown one tap away.

2Every claim carries its own citation. If the sources disagree, or there aren't enough, the answer says so.

3Bangs, Duck Player and Duck.ai show up right where they'd help.

02 · Interactive prototype

Have a go.

Open a privacy line, click a citation, type ! in the search box, or switch the answer states. Turn on the design notes to see what changed and why.

Viewing in

Answer state

Design notes

3 Search

Germany ▾Safe search: moderate ▾Any time ▾

2

Search Assist

ADsoundhaus-shop.deBased on your search, not on you

Noise cancelling headphones: free returns for 60 days

Over 40 models in stock. Shipping in 24 hours across Germany.

1Privacy line. One row per result, in plain words and backed up by colour. Open it to see trackers by category, cookie pop-ups and what DuckDuckGo will do about them.

2Answers you can check. Each sentence links to the result that backs it, and clicking a citation highlights that result. There are states for disagreement and for "not enough to go on".

3!Bang suggestions. Typing ! opens a short list of matching bangs. A result from a site that has a bang also shows a one-time tip.

4Duck Player in place. Video results say how they'll play before you click, not after.

5Opt-in filter. "Fewer trackers" hides heavy pages and always says how many it hid. Ranking stays the same.

6A friendly nudge. In other browsers the line reads "on this page", with an option to block them with the DuckDuckGo browser. It's a reason to switch based on facts, never on fear.

Prototype with made-up example sites and numbers. The ad, results and tracker counts are illustrative.

03 · Rationale

Decisions, and what I chose not to do.

1

Describe the site, never the person

Every privacy signal is a fact about a website, taken from the public Tracker Radar crawl. Nothing about the searcher is stored, inferred or sent anywhere. The feature adds no new data and keeps the existing promise.

Rejected: a personal "you were tracked X times this week" tally. It's engaging, but it means keeping a history of what someone did.

2

Plain words, not grades

"23 trackers on this page" says more than an "F", and it stays accurate as the data changes. Colour backs up the words but never replaces them, so the line works for colour-blind users and screen readers.

Rejected: letter grades on results. They feel like a verdict on a publisher, invite disputes, and tell readers less.

3

Honest AI states are a feature

When sources disagree, the answer shows both sides and says which results hold which view. When there isn't enough to go on, it says so and points to the results instead of guessing. I built the same logic in Readr, where the agent rewrites its query rather than inventing an answer.

Rejected: one confident paragraph with a "may contain errors" footer. It undermines trust exactly when people need it.

4

Friendly, never fearful

"Fewer trackers" is a chip people choose, and it always says how many results it hid. The copy states facts ("4 trackers will be blocked") rather than warnings, and the default ranking stays neutral, so publishers aren't penalised quietly.

Rejected: red warning badges and silently boosting clean sites. Scare tactics and invisible editorial decisions don't fit DuckDuckGo.

Weeks 1–2

Validate the language

Five-second tests and moderated sessions on the privacy line's wording: does "23 trackers" mean anything to people outside tech?

Weeks 3–4

Prototype with real data

Hook the prototype up to a Tracker Radar snapshot and real queries, and look for edge cases such as CDNs, first-party analytics and paywalls.

Weeks 5–8

Private A/B test

Cohorts assigned on the device, results reported as anonymous aggregate counters only, watched against the guardrail metrics.

After

Ship in stages

Privacy line, then citations, then bang tips, so each one can be measured on its own and documented in the design system with tokens.

04 · Fit

What you're looking for, and where I can show it.

The requirements from the posting next to the evidence. Where I'm still growing, I say so.

6+ years of product design, senior roles at global consumer brands

Six years of UX/UI at Appsfactory and Syndikat7 for Mercedes-Benz, Deutsche Bahn, Amazon, Samsung and tagesschau, including Head of Design / Senior UX Designer at Syndikat7. My senior experience comes from agency work rather than an in-house team, and I'd like to bring that range in-house.

Leading projects end to end, transparently, across disciplines

I designed and built Upfront Tech Jobs alone, across three interfaces: candidate, employer and moderation. I built Hane from dataset to TestFlight. At Schütte Kunstauktionen I led the company's digital transformation.

High-quality mockups, prototypes and rationale

This page: a working prototype with decisions and rejected options written out. I prototype with real code and real models, not placeholder screens.

Agentic workflows in everyday design work

Six months of full-time AI engineering training, plus Readr, a self-correcting LangGraph agent. AI agents are part of how I research, write and build every day, and this application was made the same way. The judgement and taste stay mine.

Creative briefs with goals, rationale and metrics

See section 01: insight, goals, non-goals and metrics that fit privacy. I also spent years as a journalist and editor for Freie Presse, so I write clearly and quickly.

Research and experiments, qualitative and quantitative

TestFlight feedback loops with 200+ Hane users. ML work covering EDA, threshold tuning and evaluation means I'm comfortable reading data, not just quoting it.

Being a steward of quality and supporting others

I built the MBUX Head Unit library framework in Sketch for a whole team, and the plugin I proposed for it was later adopted in several Mercedes-Benz workflows. Design systems are where I like to look after quality.

Privacy-by-design, ethically

Upfront Tech Jobs: salaries required, paid ranking always labelled, self-hosted analytics, no third-party trackers. I made those calls before I ever applied here.

05 · About me

The bridge between design and engineering.

I studied Multimedia and VR Design at Burg Giebichenstein, and Japanese Studies before that. I spent years designing apps in agencies, then trained full-time as an AI engineer. I design an interface and then build the parts behind it, which means I can test ideas with running code.

AI experiences people trustStreaming, grounding and honest error states, not just a chat window.

Prototypes with real modelsTesting ideas with running code instead of placeholder screens.

Design systems built like codeComponents, tokens and handoffs that work for engineers.

Translating between teamsGetting design, engineering and business to agree, each in their own language.

Selected clients

Mercedes-BenzDeutsche BahnAmazonSamsungtagesschauflaschenpostMercedes mePackshack

Experience

12/2025 – today
Freelance Design Engineer
Independent · UX/UI & AI
UX/UI concepts, interface design and design systems. AI/ML prototypes from data pipeline to interface.
2024 – today
Founder, Designer & Developer
Hane · hane-app.com
Native iOS app for learning Japanese: concept, UX/UI, Python data pipeline, SwiftUI, StoreKit 2. 200+ active TestFlight users.
02/2026 – 08/2026
AI Engineering (Data Science & ML)
neue fische · full-time programme
ML, deep learning and computer vision, LLMs and RAG, MLOps with Docker, FastAPI, dbt, Prefect and Grafana. Two capstones.
01/2025 – 11/2025
Auctioneer & Digital Transformation
Schütte Kunstauktionen GmbH
Led the digital transformation, managed the website and projects, ran live auctions.
05/2021 – 01/2025 · 04/2019 – 10/2020
UX/UI Designer
Appsfactory GmbH
App concepts and design for Mercedes-Benz, Deutsche Bahn, Amazon, Samsung and tagesschau. Client consulting and design management.
10/2020 – 04/2021
Head of Design / Senior UX Designer
Syndikat7 GmbH
App and website concepts and design, client consulting, design management.
2011 – today
Editor & Freelance Journalist
Freie Presse · regional daily
Articles on culture and events, editing, photography.

06 · Selected work

New products where AI meets the interface, and the work behind them.

Product designFull-stackLaunching soon

Upfront Tech Jobs

European tech job board · designed & built solo · 2026

Every listing shows its salary, every employer is verified, and paid placement is always labelled as advertising. Three interfaces: candidates browse without an account, employers land on their next decision, and moderators can't act without recording a reason in an audit log.

Why it's relevant hereTrust built into the product: required salaries, labelled ads, search that works without JavaScript, and self-hosted analytics with no third-party trackers.

Next.js · React 19 · TypeScript · PostgreSQL · Prisma · Playwright + axe

upfronttechjobs.com

iOSData pipelineUX/UIBeta

Hane

Learning Japanese on iOS · own product since 2024

Reading, vocabulary and review in one native app, with a Python pipeline that gives every text the correct furigana. The same characters can be read differently (一日 is tsuitachi or ichinichi), so the pipeline resolves that and the interface can focus on reading.

Where design meets engineeringLearners should never trip over a wrong reading, so the ambiguity gets resolved upstream. FSRS spaced repetition, on-device OCR and TTS, Live Activities.

Swift · SwiftUI · Python · fugashi · UniDic · SQLite · StoreKit 2 · Apple Vision

hane-app.com

AI engineeringInterface designCapstone 2026

Readr

Agentic RAG reading assistant for Japanese & Hebrew learners

Answers vocabulary questions and simplifies any text to a target level (JLPT N5–N1 / CEFR A1–C1), with inline annotations graded by difficulty. When retrieval comes back with low relevance, the agent rewrites the query instead of making up an answer.

Why it's relevant hereFor learners, trust is the most important part of the UX: grounded answers, token streaming so it feels fast, and difficulty you can see. It's the same thinking behind the Search Assist states above.

Python · LangGraph · LangChain · ChromaDB · Ollama · Groq · FastAPI · Pydantic · React 19

UX/UI · Design system

Mercedes-Benz "Exclusive App"

A premium app for high-end customers that follows the design system: exclusive events, merchandise, and a personal assistant you start with a speech-to-text button. Light and dark mode.

Today I'd prototype the assistant as an LLM agent with retrieval, a real model instead of a scripted dialogue.

Design system · Tooling

MBUX Head Unit Library

A Sketch library framework for every head unit component, fed by style libraries so the team could build quickly and switch styles. The plugin I proposed was later adopted in several Mercedes-Benz workflows.

Design systems are architecture: components, dependencies, central styles.

UX/UI · Web

Packshack

A landing page and pricing calculator for a fulfilment service. The old page left visitors unsure what the company did. The new one leads with the product and lets people work out their own shipping costs.

A calculator instead of a sales pitch: the product logic sits right in the interface.

07 · Skills & education

The toolkit.

Design

UX/UI conceptInterface designDesign systemsPrototypingInteraction designMotion designAccessibilityTypographyBrand & printPhotographyFigmaSketchAdobe CCAfter EffectsAffinity

Consulting

Stakeholder consultingDesign managementProject managementDigital transformationEditing

Engineering

TypeScriptJavaScriptReactNext.jsAstroTailwind CSSHTML / CSSSwift / SwiftUIPythonFastAPISQLPostgreSQLSQLiteGitDockerCI/CDPlaywright

AI & ML

RAGAgentic workflowsLangGraphLangChainPrompt engineeringEmbeddingsChromaDBOllamascikit-learnPyTorchTensorFlowComputer visionpandasPolarsMLflow

Education

2026AI Engineering (Data Science & ML)neue fische · full-time, remote
2014 – 2019BA Multimedia | VR DesignBurg Giebichenstein University of Art and Design Halle
2009 – 2014BA Japanese Studies / Applied LinguisticsMartin Luther University Halle-Wittenberg
2005 – 2006Exchange yearSagakita Senior High School, Japan

Languages

German · nativeEnglish · fluentJapanese · degree, still learning

Outside work

Board game designJapanese cultureArt exhibitionsTable footballTable tennis

08 · Why DuckDuckGo

Why I want to do this here.

Same values

When I built my own job board I chose labelled ads, no third-party trackers and published ranking rules, because I think that's how the internet should work. DuckDuckGo builds on that idea for millions of people.

Private, useful AI

Duck.ai and Search Assist raise the questions I've been working on: how to make AI answers trustworthy, checkable and optional. I'd like to work on that problem at DuckDuckGo's scale.

End-to-end ownership

I work best when I own a problem from the brief to the release. That's how all my own products came about, and it's how DuckDuckGo describes its culture.

Let's talk.

I'd love to walk you through the prototype, or start on whatever your team needs next.

sascha.thiele@pm.me LinkedInsascha-thiele GitHubsthix Leipzig, Germanyremote