The thing that trips people up about Chai Discovery is treating it like the healthcare-AI startups it gets lumped in with. It isn’t selling ambient scribing or claims automation. It trains foundation models that predict and design molecular structure: how proteins fold, how antibodies bind, whether a designed molecule will actually stick to its target in a wet lab. Co-founder Jack Dent has said every line of code in the codebase is homegrown and they aren’t taking LLMs off the shelf. That one sentence tells you what the interview is testing for.
The company was founded in 2023 by Josh Meier, Jack Dent, Matthew McPartlon, and Jacques Boitreaud, working out of OpenAI’s offices in the Mission before spinning out, per TechCrunch’s reporting. Meier co-created ESM-1, the first transformer protein language model, at Meta and was Chief AI Officer at Absci; Dent came from Stripe engineering. That pedigree sets the tone: the people interviewing you have shipped both research and production infrastructure, and they expect you to be credible on at least one of those axes and literate in the other.
Who Chai actually hires, and the two tracks
Roles cluster into two families, and which one you’re in changes almost everything about the loop. AI Research Scientist and Research roles are judged on modeling: do you understand equivariant architectures, diffusion and flow-matching for structure generation, the AlphaFold-lineage co-folding stack, and how to design an experiment that tells you whether a model got better. AI Research Engineer roles are judged on whether you can make training and inference of those models fast, correct, and reproducible: distributed training, GPU memory, data pipelines over messy structural biology data, and code you’d actually merge.
The question most infra engineers ask is whether a lack of structural-biology background is a gate for the research-engineering track. My informed read: it isn’t. Bio literacy is learnable on the job for an RE, and the founders’ own bet is that biology becomes an engineering discipline. The real gate is the modeling-and-infra bar, whether you can profile a training run, reason about a loss curve, and ship correct distributed code. A great GPU-infra engineer who has never touched a PDB file but is hungry to learn the domain clears it; someone strong on biology but shaky on the systems does not.
Most strong candidates lean hard on one track and can hold a conversation on the other. A research engineer who can’t reason about why a loss curve plateaued won’t clear the bar, and a scientist who has never profiled a training run or written a clean PyTorch module won’t either. This is the same blend you see at other frontier-model labs; if you’ve read the Sakana AI interview guide or the pages for Cohere and Mistral AI, the research-plus-engineering shape will look familiar. What’s different at Chai is the domain: the model’s output is a hypothesis a pharma partner spends real money to test in a lab, so correctness and calibration matter more than a leaderboard number.
The interview process, round by round
Chai does not publish its loop, and it’s small enough that the exact sequence flexes by role and by who’s hiring. The shape reported by candidates and consistent with how labs this size run is a fit-focused first call (often with a founder or a senior researcher, not just a recruiter), a technical screen matched to your track, one or more deep-dive rounds, and a final onsite or paid work trial. There are too few public Glassdoor data points to trust a difficulty average, so treat the table below as a realistic template, not a fixed script.
| Stage | What it is | What they’re actually checking |
|---|---|---|
| Intro / fit call (30-45 min) | Conversation with a recruiter, founder, or senior researcher about your background and why AI-for-biology | Genuine interest in the science, not just “an AI startup”; can you explain your best project crisply |
| Technical screen (60 min) | Research track: paper reasoning, ML fundamentals, structure/geometry. Engineering track: live coding, systems or PyTorch questions | Depth in your primary track and no bluffing; clean thinking out loud |
| Deep dive (60-90 min) | Walk through a project you led end to end, then extend it under questioning | Did you actually own the hard decisions; can you defend tradeoffs when pushed |
| Work trial or onsite | A scoped problem close to real work: a modeling experiment, a training-infra task, or a debugging exercise on their kind of stack | Speed and quality on realistic work; how you collaborate; whether your code or analysis is trustworthy |
| Final / references | Team fit, calibration across interviewers, offer | Consensus that you raise the bar; alignment on level and comp |
The work trial is the costliest and most decisive stage, so plan for it. Chai hasn’t published the format, but labs at this size almost always pay for a trial and scope it to a self-contained problem: a modeling experiment you run and write up, or an infra task on a representative stack, done remotely or partly onsite over a short stretch of days rather than a single afternoon. I won’t put a specific day count or rate on it because that isn’t officially stated, so ask the recruiter directly what the trial involves, how long they expect it to take, and how it’s compensated. A team that runs a serious trial will answer all three without hesitation.
Questions phrased the way they get asked
For the research track, expect open modeling questions grounded in structure prediction and generation rather than trivia. Some in the spirit of what comes up:
- Why is equivariance to rotation and translation useful for a structure model, and where would you spend the compute if you had to give some of it up?
- Chai-1 co-folds proteins with ligands and other molecules. What breaks if you naively extend a protein-only model to that setting?
- Chai-2 does zero-shot de novo antibody design at double-digit wet-lab success rates, a figure the company and FierceBiotech have cited. How would you design the eval so that number means something, given how few candidates you can physically test?
- Your validation metric improves but the pharma partner’s assay results don’t. Walk me through how you’d find the disconnect.
For the engineering track, the questions look more like a strong infra interview pointed at model training:
- A training run is GPU-underutilized at 40%. How do you diagnose it, and what are the usual suspects in a large structure-model pipeline?
- Design the data pipeline that turns raw PDB and assay data into training batches, and tell me where it silently corrupts labels.
- Here’s a slow model-serving path. Profile it and speed it up. (Reported as a hands-on exercise, not a whiteboard hypothetical.)
You’ll also get behavioral questions, but at a research lab they’re really about ownership and judgment: a time you killed your own promising idea because the evidence said so, or shipped something you weren’t proud of and what you did next. Prep these as concrete stories, not adjectives. If you want a structure for that, the STAR method holds up fine here.
Reading the comp bands without guessing
Chai doesn’t publish a leveling ladder, and I’m not going to invent numbers for you. There are two reliable ways to get the real figure. First, California’s pay-transparency law means every Chai role posted on its Ashby careers page carries a base salary range; read the band on the specific listing you’re applying to, because it varies a lot by track and seniority. Second, cross-check total compensation using levels.fyi and by asking the recruiter directly what band and level they’re targeting on the fit call, which is a normal question and tells you how serious the process is.
What you can reason about structurally: at a $3.8B valuation off a $400M round led by Index Ventures with Kleiner Perkins, Sequoia, General Catalyst, and OpenAI in the mix (per Dealroom and FierceBiotech), equity is the lever that moves an offer, not base. Ask how many shares (not just a dollar figure), the current strike or preferred price, the vesting schedule, and how they handle refreshers. A base band at a frontier lab tells you less than the equity terms do at this stage.
On location and work authorization: Chai is headquartered in San Francisco, and roles at a lab this size skew onsite or hybrid there rather than fully remote. Visa sponsorship isn’t publicly confirmed one way or the other, so if you need it, ask on the first call and check the specific posting before investing in the loop.
How to prep in the two weeks before
Read Chai’s own papers first, not summaries of them. Being able to say something specific and slightly critical about Chai-1’s co-folding approach or Chai-2’s antibody-design results is worth more than any amount of generic ML review, because it proves the interest the fit call is screening for. If you’re on the research track, refresh the AlphaFold2/AlphaFold3 lineage, equivariant and geometric deep learning, and diffusion and flow-matching for generation. If you’re on the engineering track, be fluent in PyTorch internals, distributed training (FSDP or similar), and profiling, and have a real project you can defend where you made training or inference meaningfully faster.
For both tracks, do a mock of the deep-dive round out loud, because the failure mode here isn’t lack of knowledge, it’s freezing when an interviewer pushes on a decision you made two years ago and half-forgot. This is a lab where the interviewers have built the thing you’re being asked about, so the AI-era shortcut of pattern-matching to memorized answers backfires. The broader AI-era interview playbook and the AI-native company guides hub cover how these frontier-lab loops differ from a standard big-tech loop, and the AI-startup difficulty index gives you a sense of where Chai sits relative to peers.
The candidates who do well aren’t the ones who memorized the most architectures. They’re the ones who clearly wanted to work on this specific problem, could go deep on their own past work without flinching, and were candid about the seams in their knowledge instead of papering over them. In a room of people who co-founded protein language models, that candor reads as competence.
