Clara is a senior ML engineer with no academic papers. Here is what Flow told her about this PhD, what it asked her, and the plan it made, about five weeks before it closes.
Clara is made up. The opportunity, its questions and everything Flow says are real.
Designing Responsible NLP PhD, Edinburgh
Is this for you?
Possible fitYour BSc is in a relevant subject, but the CV gives no UK 2:1 equivalent; the text also accepts extensive professional experience, and you have seven years.
Lead with this
Lead with your 2024-2025 review-fairness work: you found a language bias in a live ranking model, tested it, fixed it and spoke to the hotel owners affected.
“I co-designed the fix with small-hotel owners.”
A stretchSaferSay you interviewed 14 hotel owners before and after the change.
7 fields mapped2 ready5 need her
Three of Flow’s calls for Clara, played once. All of the plan follows, as it ran.
Part 1 · Is this for her?
Free, about two minutes. Flow read the official page and her CV.
Is this for me?
CLARA NOVAK Senior Machine Learning Engineer · Amsterdam, Netherlands clara.novak@example.com · +31 6 0000 0000 · Czech citizen SUMMARY Machine learning engineer working on ranking and language technology for seven years. I care most about the people a ranking model quietly decides for: the property owner who never learns why their listing dropped. EXPERIENCE Senior Machine Learning Engineer, Reviews & Search Ranking (Sep 2021 – present; senior since Jan 2024) Online travel marketplace, Amsterdam (about 2,000 engineers; reviews in 20 languages) - Led the review-fairness work, in a team of four engineers and one data scientist (2024-2025). - Found that our review-quality model scored machine-translated reviews lower than reviews written in the guest's own language, in 6 of 20 languages. Small independent hotels, whose guests more often write in another language, were pushed down search results as a result. - Built the test that showed it: 3,000 pairs of the same review, original and machine-translated, labelled by 12 native speakers from our customer-service teams in Lisbon, Istanbul and Kraków. - Proposed and shipped the fix: the model now scores each review in the language it was written in. About 40,000 listings were re-ranked; bookings for independent properties in the six languages rose 7% in the A/B test. - Interviewed 14 small-hotel owners with our partner-support team before and after the change. - Wrote the team's guideline for testing ranking features for language bias, now on the launch checklist of three ranking teams. Machine Learning Engineer, Search (Jul 2019 – Aug 2021) Online grocery start-up, Prague (40 people) - Built product search for Czech and Slovak queries: spelling correction and word-form matching for two languages with rich inflection. Zero-result searches fell from 11% to 4%. - First ML hire; set up the team's offline evaluation and weekly error review. EDUCATION BSc, Open Informatics (Artificial Intelligence), Czech Technical University in Prague, 2019. Graduated with honours. Thesis: "Spelling correction for Czech product-search queries" (grade A). TALKS AND WRITING - "When the translation is the bias: ranking reviews across 20 languages", PyData Amsterdam, 2025. - Internal engineering blog: "What our review model thought of Google Translate" (2024), read company-wide. - No academic publications. OTHER - Volunteer mentor, free evening Python course for career changers, Amsterdam (2023 – present). - Languages: Czech (native), Slovak (fluent), English (C2), Dutch (B1), German (A2). - Skills: Python, PyTorch, Hugging Face Transformers, learning-to-rank (LightGBM), Spark, SQL, A/B testing, annotation design.
Check this first: At least a 2:1 undergraduate and/or Master's degree in a relevant subject, though extensive professional experience can also count.
They say: “We expect successful applicants will have at least a 2:1 (upper second class) Undergraduate and/or a Master’s degree in a relevant subject”
You shared: “BSc, Open Informatics (Artificial Intelligence), Czech Technical University in Prague, 2019. Graduated with honours.”
Your BSc is in a relevant subject, but the CV gives no UK 2:1 equivalent; the text also accepts extensive professional experience, and you have seven years.
They say: “we are looking for applicants that clearly align with our five Skills Domains”
You shared: “Found that our review-quality model scored machine-translated reviews lower than reviews written in the guest's own”
Your finding of language bias in a deployed model is the core of the Responsible NLP data & models domain.
They say: “any projects you have worked on which demonstrate some element of research skills”
You shared: “Built the test that showed it: 3,000 pairs of the same review, original and machine-translated, labelled by 12”
Your paired test with 12 annotators is a real method, but it is industry work and you have no academic publications.
They say: “We’d expect some awareness of recent debates and research of relevance to the focus of the CDT”
Not in what you shared
Nothing on the CV cites research papers or venues like FAccT or ACL, so the selectors cannot see this yet.
They say: “we will be looking out for an interest and desire to collaborate with others at application time”
You shared: “Interviewed 14 small-hotel owners with our partner-support team before and after the change.”
You worked with support staff, annotators in three countries and affected hotel owners, which shows cross-team work with real people.
They say: “we’re looking for students that have a long-term ambition to translate the research from their PhD”
You shared: “Wrote the team's guideline for testing ranking features for language bias, now on the launch checklist of three”
Your guideline shows findings turned into practice, but the CV does not state a longer-term ambition.
Lead with your 2024-2025 review-fairness work: you found a language bias in a live ranking model, tested it, fixed it and spoke to the hotel owners affected. It matches the Responsible NLP data & models domain and the CDT's focus on work in-the-world.
You shared: “Led the review-fairness work, in a team of four engineers and one data scientist (2024-2025).”
Short questions, only for what her CV does not say. She skips what she cannot answer yet, and the plan leaves it open.
What did the 12 labellers judge in each review pair?
Whether the translated review says the same thing as the original, and whether it would change how they rate the hotel
How did you handle disagreement between annotators?
Majority vote
What other volunteer or advocacy work on responsible AI have you done?
None beyond mentoring on the evening Python course
Which two or three ACL, FAccT or AIES papers on multilingual bias have you read?
Clara skipped this
Do you want to name a third domain?
No
Which CDT partner organisations or supervisors would you like to work with?
Clara skipped this
Part 2 · The plan
Flow’s summary
You are a possible fit: you have a real, measured bias finding in multilingual NLP, but no academic papers and no sign yet of reading the responsible-NLP literature. Stage 1 is judged mainly on fit to the five Skills Domains, so the Statement of Alignment matters most. Start by reading two or three recent ACL or FAccT papers on translation or multilingual bias, so you can cite them there. Then write the research experience answer around your test design.
Lines she might write, read the way a selector reads them.
“My test showed that machine-translated reviews were scored lower in 6 of 20 languages.”
What you shared backs thisYou shared: “Found that our review-quality model scored machine-translated reviews lower than reviews written in the guest's own”
“I have published research on bias in NLP.”
Not in what you sharedSafer: Say you gave a 2025 PyData talk and wrote a 2024 internal blog post on translation bias. The CV states: No academic publications.
“My fix made search fair for small independent hotels.”
A stretchYou shared: “bookings for independent properties in the six languages rose 7% in the A/B test.”
Safer: Say that scoring each review in its own language re-ranked about 40,000 listings, and bookings for independent properties in the six languages rose 7% in the A/B test.
“I co-designed the fix with small-hotel owners.”
A stretchYou shared: “Interviewed 14 small-hotel owners with our partner-support team before and after the change.”
Safer: Say you interviewed 14 hotel owners before and after the change. Say it was interviews, not co-design, unless you can confirm owners shaped the decisions.
“I led the team's review-fairness work.”
What you shared backs thisYou shared: “Led the review-fairness work, in a team of four engineers and one data scientist (2024-2025).”
The fact
Test: 3,000 review pairs, labelled by 12 native speakers
Leads“Research experience (within and beyond academia)”
The form asks for methods and research skills, and a controlled paired test with native-speaker labels shows method better than a job title.
The fact
Model scored translated reviews lower in 6 of 20 languages
Leads“Statement of Alignment”
It is your own prior work on bias in data and models, so it grounds the main-domain choice and your research direction.
The fact
Shipped fix: 40,000 listings re-ranked, independent bookings up 7%
Leads“Professional or other relevant experience”
It shows you took a finding into a live product with a measured result, which is the translating-research-into-practice trait the CDT lists.
Left out, and why:
Research experience: methods and projects that show research skills
Describe the question, the paired-review test design, the 12 native-speaker labellers and the 6-of-20-languages result. Add [confirm: how you measured the score gap] from your own records.
From her CV: “Built the test that showed it: 3,000 pairs of the same review, original and machine-translated, labelled by 12”
Professional experience and how it sets you up for the CDT
Lead with the shipped fix and its measured result. Then add the guideline that three teams now use, to show you apply responsibility in practice.
From her CV: “Proposed and shipped the fix: the model now scores each review in the language it was written in.”
Statement of Alignment: main domain, research directions, second domain, interdisciplinary work
Name Responsible NLP data & models as your main domain, built on your translation-bias finding. Tie the hotel-owner interviews to a second domain such as Co-creation and NLP futures, and to working with other disciplines and partners.
From her CV: “Interviewed 14 small-hotel owners with our partner-support team before and after the change.”
Publications, outputs and links (optional)
Link the 2025 PyData talk and the 2024 blog post. Check that the blog is shareable outside your company before you link it.
From her CV: “"When the translation is the bias: ranking reviews across 20 languages", PyData Amsterdam, 2025.”
Awareness of recent responsible-NLP research
The CV shows none yet. Read two or three recent ACL or FAccT papers on multilingual or translation bias. Cite them in the Statement and say where your finding agrees or differs.
Prior degrees (2:1-level, or strong professional experience)
Enter the degree as it appears on your transcript. Do not convert the grade yourself. Your seven years of ranking work also supports the experience route.
From her CV: “BSc, Open Informatics (Artificial Intelligence), Czech Technical University in Prague, 2019. Graduated with honours.”
Marked: from her answers to Flow’s questions[confirm: …]: a fact only she can add
Research experience (within and beyond academia)
What this judges: Whether you can design a controlled test and read its result
From her CV: “Built the test that showed it: 3,000 pairs of the same review, original and machine-translated, labelled by 12”
From her CV: “Led the review-fairness work, in a team of four engineers and one data scientist (2024-2025).”
From her CV: “No academic publications.”
My research experience comes from industry, not from a degree or a journal. In 2024-2025 I led the review-fairness work on the Reviews & Search Ranking team of an online travel marketplace, with four engineers and one data scientist. The question was whether our review-quality model scores a review differently once it has been machine-translated.
To answer it I built a paired test: 3,000 pairs of the same review, once in the original and once machine-translated, labelled by 12 native speakers from our customer-service teams in Lisbon, Istanbul and Kraków. For each pair, the labellers judged whether the translated review says the same thing as the original, and whether it would change how they rate the hotel. Where labellers disagreed, I used the majority vote. Because both sides of a pair carry the same content, a difference in the model's score points to the translation and not to the review. [confirm: how you measured the score gap between the two sides of a pair]. The test showed that translated reviews were scored lower in several of our languages.
This work taught me to design annotation, to keep a comparison controlled, and to check a result against a live A/B test before trusting it. I have no academic publications. I would like the CDT to give me the formal training to take this kind of question further. [confirm: any link to the test write-up you can share outside your company]
It leads with the paired test because this question asks for methods, and the test shows method more clearly than a job title.
Professional or other relevant experience
What this judges: Whether you take responsible-AI findings into real products and teams
From her CV: “Proposed and shipped the fix: the model now scores each review in the language it was written in. About 40,000”
From her CV: “Wrote the team's guideline for testing ranking features for language bias, now on the launch checklist of three”
From her CV: “First ML hire; set up the team's offline evaluation and weekly error review.”
For seven years I have worked on ranking and language technology, and the work I want to build on is the most recent. As a senior machine learning engineer at an online travel marketplace, I proposed and shipped a fix for a language bias in our review-quality model: the model now scores each review in the language it was written in. About 40,000 listings were re-ranked. In the A/B test, bookings for independent properties in the affected languages rose 7%. That result matters to me because the people who gained were small independent hotels, who had no way to see why they had been pushed down.
I then wrote the team's guideline for testing ranking features for language bias. It is now on the launch checklist of three ranking teams, so the check happens before launch and does not depend on one person remembering it. This is the gap the CDT describes between ethical principles and practice, and I have closed a small part of it in a real product.
Before this, I was the first ML hire at a 40-person grocery start-up in Prague. There I set up the team's offline evaluation and weekly error review, so I know how to work in a small team with few resources.
What I lack is a research setting where I can study these problems with other disciplines and for longer than a product cycle allows. That is what I would like the PhD to give me.
It leads with the shipped fix and its measured result because this section asks how your background prepares you to put research into practice.
Statement of Alignment
What this judges: Whether your interests fit the domains, the field, and cross-disciplinary work
From her CV: “Found that our review-quality model scored machine-translated reviews lower than reviews written in the guest's own”
From her CV: “Interviewed 14 small-hotel owners with our partner-support team before and after the change.”
From her CV: “Small independent hotels, whose guests more often write in another language, were”
Main domain: Responsible NLP data & models. My interest comes from a result I found at work. Our review-quality model scored machine-translated reviews lower than reviews written in the guest's own language, in 6 of our 20 languages. Small independent hotels, whose guests more often write in another language, were pushed down search results as a result. This is a bias that comes from how data and models are built, and it affects people who never see the model. I want to develop deeper expertise in locating this kind of bias and measuring it across languages.
Research directions. I would like to ask how bias against translated or non-native text arises in the data and models behind ranking and quality systems, and how to test for it before deployment. I would also like to ask which languages and cultures this affects most, and how far it holds outside English-centred settings, which the CDT lists as part of this domain. Another question is how teams can document and share such tests so that they are reproducible. [confirm: cite two or three recent ACL, FAccT or AIES papers on bias in multilingual or machine-translated text that you have read, and say where your finding agrees with them or differs]
Second domain: Co-creation and NLP futures. When I interviewed 14 small-hotel owners with our partner-support team, before and after our change, I learned how much I had not known about how they experience a ranking system. I want to learn methods that let the people affected by a system help define what counts as a problem, and not only assess it after the fact. I am not naming a third domain.
Working across disciplines and with partners. My bias test relied on native speakers from customer-service teams in three countries, and my interviews relied on a partner-support team. In both cases I needed people whose knowledge I did not have. I would bring the same habit to students from law, design or linguistics, and I would like to learn from them how they would frame the problem. With industry partners, I can bring seven years of experience of shipping ranking systems. [confirm: any partner organisations or supervisors from the CDT pages you would like to work with]
It leads with the 6-of-20-languages finding because it is your own prior work on bias in data and models, which grounds your choice of main domain.
2 ready from her CV5 need her facts
Basic personal information
Needs youName, email and phone from your CV. Fee status (Home or Overseas) is your choice to confirm from UKRI rules; check it.
From her CV: “clara.novak@example.com · +31 6 0000 0000 · Czech citizen”
Information about your prior degrees
Needs youBSc, CTU Prague, 2019, honours. The CV gives no grade on a UK scale, so take the exact grade from your transcript.
From her CV: “BSc, Open Informatics (Artificial Intelligence), Czech Technical University in Prague, 2019. Graduated with honours.”
Research experience (within and beyond academia)
Ready from what you sharedThe translation-bias test: design, 3,000 pairs, 12 labellers, finding in 6 of 20 languages.
From her CV: “Built the test that showed it: 3,000 pairs of the same review, original and machine-translated, labelled by 12”
Publications, outputs and links (optional)
Needs youLinks to the 2025 PyData talk and the 2024 blog post. The CV has no URLs, so add them yourself.
From her CV: “"When the translation is the bias: ranking reviews across 20 languages", PyData Amsterdam, 2025.”
Professional or other relevant experience
Ready from what you sharedReview-fairness lead role, the shipped fix with its A/B result, and the guideline now used by three teams.
From her CV: “Led the review-fairness work, in a team of four engineers and one data scientist (2024-2025).”
Statement of Alignment
Needs youCover all four points: main domain, research directions, second domain, working with other disciplines and industry. Your reading and views on future directions are not in the CV, so you must add them.
From her CV: “Interviewed 14 small-hotel owners with our partner-support team before and after the change.”
Potential supervisors (optional)
Needs youOptional. Only you can choose names, from the CDT supervisor webpage. The CV is silent.
No sign of recent responsible-NLP research in your CV
Read two or three recent ACL or FAccT papers on bias in multilingual or machine-translated text. Note how your finding agrees or differs, and cite them in the Statement of Alignment.
this weekHow you measured the score gap, and the test details, are not in the CV
Pull the metric, the comparison method and any significance check from your team's records. Check what you may share outside the company.
before you write the research experience answerFee status (Home or Overseas) and your grade equivalent are not in the CV
Check the UKRI fee-status rules for a Czech citizen living in the Netherlands, and ask CTU for your transcript and grade explanation. Do not guess either.
before you submitNo second domain or supervisor chosen yet
Read the five domains and the CDT supervisor webpage. Pick a second domain you can back with the hotel-owner interviews, and decide whether to list any supervisors (optional).
this weekPart 3 · Before she sends
The fit check is free. The full plan is INR 499 or USD 9, refunded within 7 days if it doesn’t help. A real plan is built from what you share and the real form, and a person reads it once it opens.
Flow’s output as it ran on 28 Oct 2026, unedited. Clara and her CV are made up.