Why You Fit, and What You Already Promised
The project in its own words, your CV mapped onto what it actually needs, and — most importantly — what your cover letter already committed you to. Read this with the letter open beside you.
Listen
0. What the project says it is, in its own words
Before the fit argument, get the project's self-description exactly right, because two phrases in it should change how you answer.
- ATLANTIS is a five-year project on computational antitrust, led by Schrepel at VU Amsterdam, funded by an ERC Consolidator Grant. Its own one-line summary: it studies how computation changes antitrust law, and builds a robust legal regime for it.
- The central question, stated by VU: not *whether* these tools will be used, but how they can be used in ways that respect fundamental rights and procedural guarantees. Learn this sentence. It tells you the project's premise is settled — agencies are going to compute — and the research question is the conditions.
- Schrepel's own framing of the tension: agencies can only act effectively if they can observe and understand markets accurately; in digital environments that is becoming impossible without computational techniques; and at the same time these tools must always operate within the rule of law and respect the procedural safeguards owed to companies and individuals.
- ATLANTIS combines three things: legal analysis, computational science, and institutional economics.
- And the line that matters most for you: researchers trained in computational and economic methods will join the VU Amsterdam Faculty of Law — described as an uncommon combination in the field and a first for the faculty.
- **The agency network is *more than 75* antitrust agencies worldwide — that is VU's own figure. And crucially: those networks give the team access to practices and challenges that are often not visible in public documents.**
- Outputs: a legal and institutional framework, plus numerous empirical studies, workshops and publicly accessible documentation.
- Applications closed 15 September 2026. You submitted on 14 September — one day before.
1. The fit argument, mapped onto what the project actually needs
The project needs three things from this position. Here is each, with what you have, honestly graded.
| What the project needs | What you bring | Strength |
|---|---|---|
| Someone who can build the measurement — empirical studies that the legal strands can rely on | An evaluation harness with a paired three-condition design, exact statistical tests, a parser-scored probe and per-run provenance logging. A 20-year district panel with DAGs, do-calculus, hierarchical models and a Bayesian counterfactual. Two production systems under real cost and rate limits | Strong, and genuinely demonstrated rather than claimed |
| Someone who can translate between the two sides — the stated reason the position exists | Three degrees: computer applications, law, data science. You can read a statute and a model. This is the barrier most candidates cannot cross, and it is why the position is described as an uncommon combination | Your single strongest card. Lead with it |
| Someone who will function inside a law faculty — remembering this is *a first for the faculty* | VU summer school in formal logic, a full LL.B., published legal writing, and procedural and evidence law training. But you have mostly worked alone, and that is the honest gap | Adequate, and the thing to address head-on rather than hope they miss |
| Institutional economics — named as the third pillar | You do not have this. You have applied econometrics — mixed-effects models, fixed-effects panels, causal inference — which is not the same thing | A real gap. Name it before they do |
2. What your cover letter already committed you to — read this twice
Everything else your letter asserted, and whether you can now defend it:
| What the letter claims | Can you defend it? | What to add or correct |
|---|---|---|
| The Mens Rea paper was written during a hackathon sprint run by Apart Research | Yes — and use it. It explains the sample size | *This is the best available answer to 'why only fifteen runs'*: it was a time-boxed sprint, so the design had to be something defensible at small n, which is why it is paired and uses exact tests rather than approximations. Then add that you found and fixed a v1 error afterwards — so the sprint produced the design and the correction came from your own later audit |
| Modelled actus reus and mens rea with POS tagging, word embeddings and First Order Logic | Yes, fully. Full text read | Add the finding, which is the valuable part: the Bi-LSTM hit 86.89 percent validation and still called *she was struck with joy* murder. FOL caught it because it checks elements cumulatively. And the limitation: hand-built rules do not transfer across offence types — which is the argument for neurosymbolic rather than symbolic |
| Built a classifier over the 2023 Indian criminal code using TF-IDF, GloVe and WordNet, returning punishment, cognizability, bailability, jurisdiction | Yes. 442 provisions, 336 of 400 benchmark sentences, 84 percent | Volunteer the caveat first: you wrote the benchmark sentences, so test set and author share a vocabulary and 84 percent is optimistic. And the burglary-not-theft example in your letter is a good one — it is exactly why element-checking beats similarity scoring |
| Astroformer: 28,000 monthly actives, RAG pipelines, output guardrails, live testing, Next.js and FastAPI | Yes, verified against the repository | You can go several questions deep: three hosted LLM providers with failover, an RPM guard, Postgres, Cloudflare R2, JWT and bcrypt, hashed OTPs. And the architectural point: planetary positions come from Swiss Ephemeris, not the model — ground-and-generate with a verified numeric source |
| LawReformer: multilingual support in six Indian languages | Check this. The public site presents India, UK and US jurisdictions and six tools; the six-language claim is the one I could not verify from the site | Be ready to describe it accurately rather than as written. If the multilingual support is on ai.lawreformer.com specifically, say so. Do not restate a feature claim you cannot demonstrate |
| Outlier AI: evaluated model outputs across law, programming and linguistics, documenting failures using rubrics | Yes | Strengthen it: that is preference ranking and rubric-based scoring, which is the human side of RLHF. You supplied the data that trains reward models, and it is why your own harness votes over three judge calls and records agreement |
| The crimes paper applied DAGs, Bayesian Belief Networks, do-calculus and BSTS to 22 years of data — described as *ongoing* | Yes — and you have an update | *It is no longer ongoing.* It was posted as a SocArXiv preprint on 19 September, five days after you submitted. Lead with that as news. Headline results: Moran's I 0.095 at p 0.151, ICC 0.6 percent, urbanisation significant at p 0.004, and +88,879 reported cases a year after the 2013 amendment |
| *I have written articles for legal journals and blogs previously* | Yes — one journal article | The IoT cyber-security piece, Indian Journal of Legal Research, 2022, second author with Upasana Ghosh. Comparative EU/US/India. Be precise that it is one article and you were second author |
| *I want the depth, and honestly I want to do this properly, with a team, instead of figuring it out alone* | This is the best sentence in your letter | Repeat it almost verbatim if fit comes up. It is true, it is modest, and it answers *why a PhD rather than continuing to build* without any posturing |
3. Project A and Project B — you named them, so be ready
Your letter addressed the two legal-track projects by letter. That means the panel may well pick up your own framing, so have a position on each.
| What it is | What you said you would bring | Sharpen it to this | |
|---|---|---|---|
| Project A — the data problem | Documenting and interrogating the data agencies actually hold | Experience with dirty data; care about distinguishing a true signal from an artefact of collection; *finding out what is left out, what is unofficial, and what is inconsistent*; evidence and procedure training | *This is your strongest project fit and your letter already makes the argument.* Sharpen it with the concrete transfer: police records measure reporting, not offending; a cartel screen measures detection, not collusion. Your DAG plus counterfactual time series toolkit separates an administrative shock from a real change — and the 2013 amendment is to crime statistics what deploying a screen is to enforcement statistics |
| Project B — the fairness problem | Auditing algorithmic enforcement tools for bias | Self-taught practical experience from building two live systems with guardrails and live failure-mode testing; rubric-based model evaluation at Outlier | Upgrade this considerably. Guardrails on a consumer product is weak for an audit project. The strong version: fairness is not one property. Demographic parity, equalised odds and calibration cannot generally all hold at once unless base rates are equal across groups, which they are not. So choosing between fairness criteria is a normative decision currently made by a default parameter, and the contribution is making that choice visible rather than claiming to dissolve it. Then add the stability test: refit on a slightly different sample and see whether the *same firms* are flagged — because if the flagged set moves, who gets investigated is partly arbitrary, and that is a legal defect rather than a tuning issue |
4. Questions your letter specifically invites
| Likely question | Say this |
|---|---|
| *You say you built LawReformer because charging aggrieved people felt wrong. How is that going?* | *Very few users, and I think that is the informative part. It is free, so cost was not the barrier. People do not go looking for legal information until they are already in trouble, and by then they want someone to act rather than a tool to read. So access to justice is not primarily an information-supply problem, and I had assumed it was. That is a finding about my own assumption rather than about the software* |
| *You applied one day before the deadline.* | Do not over-explain. *I found the post late and I would rather have had longer, but I did not want to miss it.* Then move on. Nobody cares unless you make it a topic |
| *Your letter says you need the right research environment. What specifically?* | *Three things. Someone to tell me when a measurement will not bear the argument being built on it — that is the one I most lack working alone. Access to the legal tracks' actual questions early, so I build what is needed rather than what is interesting. And the data, because the questions I find most important are access-constrained rather than method-constrained* |
| *Why should we take a candidate with no EU law background over a European competition lawyer who codes?* | Answer it directly rather than deflecting. *If you can find one, that may well be the better hire. My argument is that the combination is rarer than it looks — a lawyer who codes usually codes at the level of a notebook, and the position needs someone who has run a live system, designed an evaluation and found their own error in it. And the doctrine is learnable in a way that the instinct to distrust your own measurement is not. I have spent three weeks on the instruments since I applied; I have spent three years learning to doubt my own numbers* |
| *Your work is all Indian law. Does any of it transfer?* | *The doctrine does not, and I would not pretend otherwise. The structure does. My crimes paper is about police records measuring reporting rather than offending — a measurement problem wearing the clothes of a behavioural finding. That is the same structure as a screen measuring detection rather than collusion. And the Bharatiya Nyaya Sanhita work was about mapping an old code onto a new one, which is a renumbering and domain-shift problem rather than an Indian one* |
| *You mention the VU summer school. What did you actually take from it?* | *Logic as a Tool for Modelling, in 2024, on a scholarship — first and second-order logic and formalising real scenarios. What I took from it is the thing I keep coming back to: formalising a provision forces you to choose a reading. The quantifier order in *every firm agreed with some firm* against *there is one firm every firm agreed with* is the difference between scattered bilateral deals and a hub-and-spoke cartel — same symbols, reordered. English leaves that ambiguous and the formalisation cannot. That is partly why I think a symbolic layer is the interesting direction, and it is why I was already at VU two years before this vacancy existed* |
5. Collaboration and translation — the competency with the most evidence behind it
In a fit interview this is the heaviest-weighted thing in the room, because the project is three strands that have to talk to each other and you would be among the first computational researchers in a law faculty. And the standard mistake is to answer it with enthusiasm — *I love working with people from different backgrounds* — which is unfalsifiable and therefore worth nothing. You can answer it with evidence instead, and almost nobody can.
| The work | The two fields it joins | What it proves about working across them |
|---|---|---|
| Mens Rea Evaluator | Criminal-law doctrine × LLM behavioural evaluation | You took a legal concept and turned it into a testable experimental condition. That is the exact operation the project needs — and you then corrected the framing yourself when the concept did not transplant |
| Beyond the Dark Figure | Criminology and criminal procedure × causal inference | You read a statutory amendment as an intervention and measured it. Legal change as an identification strategy |
| A Hybrid Approach to Legal Text Summarization | Legal reasoning × NLP | You discovered empirically that neither statistical nor symbolic alone is sufficient for legal text. A negative result reached by actually trying both |
| IoT cyber-security | Comparative EU/US/India law × device security | Written with a co-author, as second author, as a second-year law student. Your only genuinely collaborative publication — use it when asked about co-working |
| DiagnoChat | Clinical triage × classical machine learning | A three-person team where you were second author. And you can now read your own student project back through the AI Act and the Medical Device Regulation, which is translation in the harder direction |
| Master's thesis | Pedagogy and curriculum × LLM engineering | Ten-stage system, two supervisors — one from mathematics, one external. Four months delivering across both |
| LawReformer | Indian legal procedure × rule-based systems | Shipped. And its failure taught you something about users rather than about code |
| Astroformer | A consumer domain × production engineering | Payments, storage, auth, rate limits, a non-technical back office. Evidence you can build for people who are not you |
Translation runs both ways, and only one direction is usually prepared for
6. Corrections to carry into the room
- **The agency network is *more than 75*** — VU's own figure. Not 80-plus. Or just say *most agencies worldwide*, which is his own phrasing.
- Do not argue against agencies using computational tools. The project's premise is that they will; the question is how. Frame every criticism as a condition of proper use.
- He already says safeguards are owed to companies as well as individuals. Your undertakings argument is his premise, not your uphill battle.
- You would be among the first computational researchers in that law faculty. The risk they are managing is not whether you can code — it is whether you can work inside a law faculty. Address that, not your technical credibility.
- Institutional economics is the third named pillar and you do not have it. Applied econometrics is not the same thing. Name the gap.
- **The crimes paper is no longer *ongoing* — it was posted 19 September, five days after you applied. Lead with it as an update.**
- The Mens Rea paper came out of an Apart Research sprint — that is your legitimate answer on sample size, and the v1 fix shows you kept working on it after the sprint ended.
- Check the six-languages claim about LawReformer before Wednesday. Describe what the site actually does.
- The thesis needs three volunteered corrections: the fine-tune ran on rented GPUs (Colab and RunPod) with development on the MacBook — not the MacBook alone, as the write-up says in places; Simple English Wikipedia not English Wikipedia; and training answers and reference answers both generated by GPT-4o Mini, so the overlap metrics measure imitation rather than accuracy. Full detail in the thesis chapter.
- And revoke the API key on page 37 of the thesis if you have not already.