REVISION 2 — Your Cover Letter
Your actual letter, paragraph by paragraph: every claim you made, what each one invites them to ask, the four things that have changed since you sent it, and the two claims that need care. This is the document in front of them.
Listen
0. Why this is the chapter to read first
1. The story you told — and the one you left out
Your letter gives a four-step account of how you got here. Know it, because any *why this path* question will be read against it.
The letter's own narrative
2. Every claim you made about your own work
| What your letter claims | Be ready for | The sharpened version |
|---|---|---|
| *The Mens Rea Evaluator* — written during an Apart Research hackathon sprint, to study *whether an AI model reasoning about culpability actually follows existing legal precedents, or it quietly drifts from it when a false precedent is injected into it through RAG, and whether or not it reveals the false premise when interrogated* | The most likely question in the interview, because this is the most ATLANTIS-shaped thing you described. And *is the sample not very small?* | Your description is accurate — keep it. Add the numbers: *zero of fifteen in the clean condition, nine of fifteen with partial manipulation, fifteen of fifteen when poisoned.* Then the finding: zero of fifteen showed any cognizable intent, while the model's account of its own decision stayed fluent and plausible. **Then the correction, unprompted: *mens rea does not transplant to AI, and the paper is not arguing that it does — what the result bears on is the duty to give reasons*** |
| **Modelling *actus reus* and *mens rea* using POS tagging, word embeddings and First Order Logic**, *such that an offense can be classified by following formal methods of logical reasoning rather than having only a probability associated with it based on vectors* | *Why does formal logic help?* — and this is a question you can answer better than almost anyone | Use the quantifier example from the VU course. *Every firm agreed with some firm* against *there is one firm that every firm agreed with* — same words, reordered, and it is the difference between scattered bilateral deals and a hub-and-spoke cartel. *English leaves that ambiguous; the formalisation cannot. That is why I think the inference step should be symbolic even if the extraction step is learned* |
| The offence classifier under the 2023 Indian criminal code, using TF-IDF, GloVe and WordNet — with your own worked example: *he broke into my house and stole jewellery* maps to burglary, not merely theft, returning punishment, cognizability, bailability and jurisdiction | This is the best concrete example in your whole letter and they may well ask about it. *How does it decide burglary rather than theft?* | Answer with the legal structure, not the technology. *Because theft plus unlawful entry is a different offence from theft, so the classifier has to detect the entry element and not just the taking. **That is exactly where similarity alone fails — *broke into* and *walked into* are close in embedding space and legally decisive.** It is the clearest case I have for why a symbolic layer has to sit on top of a statistical one* |
| Astroformer — *a full stack product used by over 28,000 active users every month*, where building Astro Chat and Compatibility Chat required *RAG pipelines, output guardrails for safety, rigorous live testing to catch failure modes*, on Next.js and FastAPI | *How is it deployed?* and *what do you mean by guardrails?* | Deployment: Vercel for the frontend, Render running a Docker container for the Python API, no GPUs — every model call goes to a hosted provider. And lead with the architectural idea rather than the subject matter: *the chart is computed deterministically with an ephemeris library and injected into the prompt with an instruction to use only that data, so the model does prose and never arithmetic. Ground and generate — the legal version is identical: never let the model invent the input* |
| LawReformer — *a smaller legal tech product with multilingual support in six Indian languages*, teaching *similar lessons about keeping a live system honest* | Check this claim tonight — see section 4 | And the strong point: it is rule-based by deliberate choice. *A generative model that invents a procedural deadline for someone with no lawyer does active harm. A rule-based tool can only output what I wrote, which is checkable and wrong in predictable ways rather than unpredictable ones* |
| Outlier AI — evaluating model outputs *across domains like law, programming, and linguistics*, giving *practice noticing where models fail and documenting such issues clearly using rubrics* | *What did you learn from it?* — likely from Goanta, because this is her method | *What that work teaches you is how much the rubric does. Two careful people with the same instructions disagree far more than you expect, and the disagreement is nearly always a gap in the rubric rather than a lapse of attention. So a coding scheme has to be written and tested before labelling starts, not after* |
| The crime paper — DAGs, Bayesian Belief Networks, Do-Calculus and Bayesian Structural Time Series on 22 years of Indian crime and census data, to *identify the demographic factors driving violence and measure a crime reporting shock tied to a legal amendment from the year 2013* | **High. And *what did you find?*** and *what are its weaknesses?* | Lead with the update: it is published now. Then: *no spatial clustering, Moran's I 0.095 at p 0.151; state policy explains under one percent of local variance; and the 2013 amendment produced about 88,879 additional recorded cases a year against the counterfactual.* **Then volunteer the hole: the literacy gap is not statistically significant, and my abstract uses the word *proving* about its causal effect. Identification and precision are separate questions** |
3. Project A and Project B — in your own framing
You addressed both legal-track projects directly, so both are fair game. Here is what you offered for each, and where it needs strengthening.
What you promised against what will hold
4. The two claims that need care
5. The questions your letter specifically invites
| Likely question | Say this |
|---|---|
| *You say you are not familiar with EU competition law.* | You wrote: *it would be untrue for me to say that EU competition law is an area that I am very familiar with... I will need to learn about GDPR and the EU AI Act correctly, which I am willing to do.* And they interviewed you anyway, so the gap is priced in. The follow-through is the whole answer: *When I wrote to you I said that, and I have spent the time since doing exactly that.* Then give one specific thing you have learned, so it is demonstrated rather than asserted |
| *You describe this as a support role. Is that how you see it?* | You committed to that reading in writing, so own it and then ask the real question. *That was my assumption from the vacancy text, and it is how I would be happy to work — the measurements should serve the legal tracks' questions rather than my own interests. What I am less sure about is where the independent contribution sits, since a funded PhD still has to produce a dissertation. I would want to understand how those two fit together* |
| *You say you cannot go further without the right research environment. What specifically?* | You wrote: *I want the depth, and honestly I want to do this properly, with a team, instead of figuring it out alone without much guidance.* Make it concrete, three things: *Someone to tell me when a measurement will not bear the argument being built on it — the one I most lack working alone. The legal tracks' questions early, so I build what is needed rather than what interests me. And the data, because the questions I find most important are access-constrained rather than method-constrained* |
| *What is your biggest weakness?* | One answer, and your own letter already set it up. *I have worked alone. The consequence is that nobody has ever told me a measurement would not bear the argument I was building on it, so I have had to be my own referee — and that is a worse system than having one* |
| *Why should we take you over a competition lawyer who codes?* | Answer directly; do not deflect. *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 this position needs someone who has run a live system, designed an evaluation and found their own error in it. And doctrine is learnable in a way that the instinct to distrust your own measurement is not* |
| *How is LawReformer going?* | Honest, and the honesty is the content. *Very few users, and that is the informative part. It is free, so cost was not the barrier. People do not look 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* |
| *Your legal education is Indian. Does any of it transfer?* | *The doctrine does not and I would not pretend otherwise. The structure does. And the part that transfers most directly is the one I named in my letter — evidence and procedure. That training is about asking what a given piece of material is capable of proving, which is exactly the question to ask a screen* |
| *You mention the VU summer school. What did you take from it?* | **You wrote that it is *partly why I trust myself with formal reasoning of this kind*, so back that up.** *Logic as a Tool for Modelling, 2024, on a scholarship — first and second-order logic and formalising real scenarios. What I took from it is that formalising a provision forces you to choose a reading, and the quantifier-order example is the one I keep returning to* |