Most AI-lab interviews eventually ask some version of “how would you make this bigger.” Decart’s ask the opposite: how do you make a generative model finish a frame in the ~33 to 50 milliseconds you get at 20 to 30 fps, over and over, without the picture melting after a few seconds. That single latency budget shapes almost everything they screen for, and it’s why a strong LLM-training background alone won’t carry you here.
Decart was founded in 2023 by Dean Leitersdorf and Moshe Shalev, with a San Francisco R&D center led by Kfir Aberman, a co-author of the DreamBooth paper who came from Snap and Google Research. The company is best known for two products. Oasis (November 2024) is an interactive world model the company said passed a million users within three days. MirageLSD, released the following summer, is a low-latency video-transformation model that restyles a live stream frame by frame. Press in mid-2026 put the team around 50 people, an order-of-magnitude figure rather than a hard count. Underneath both sits the part the engineering interview cares about: a hand-written inference stack in C++ and CUDA that the company claims beats general-purpose engines like vLLM and TGI on its workloads, plus a portability layer (they call it DOS) meant to run models across different accelerators.
What they’re actually screening for
The through-line is real-time systems thinking applied to generative models. Plenty of ML engineers can fine-tune a diffusion model; far fewer can tell you where the 40 milliseconds went. Decart wants people who reach for a profiler before a bigger model, know whether a kernel is memory-bandwidth-bound or compute-bound, and treat a dropped frame as a bug rather than a rounding error.
Three buckets show up repeatedly in the roles. Low-latency inference: kernel fusion, quantization, KV-cache handling for streaming, batching that doesn’t blow the per-frame deadline. Generative-video research: diffusion and autoregressive video models, distillation down to few-step or single-step sampling so a model can run live, and the failure modes that appear only over long horizons. And the systems glue: streaming pipelines, GPU memory management, multi-chip portability, and the unglamorous work of keeping a model warm and fed at 30 fps.
If you’ve interviewed at other inference-heavy shops, the closest analogues on the prep spectrum are places like Modal Labs, Fireworks AI, and Together AI. Decart leans harder into video and real-time constraints than any of them, so the systems bar feels similar but the domain questions are more specialized.
A reconstructed interview loop
Nobody has posted a public Decart interview writeup yet, so the table below is reconstructed from the company’s job descriptions and the way comparable real-time-inference labs run their loops. Treat it as a reasonable map, not a leaked rubric. Round names and counts vary by role and by whether you’re interviewing for the Tel Aviv, northern-Israel, San Francisco, or New York team.
| Decart interview stage | Focus | What Decart is checking |
|---|---|---|
| Recruiter / hiring-manager screen | Background, why real-time video, latency intuition | Can you reason about a frame budget out loud, rather than only talk models |
| Technical phone screen | Coding or a systems problem, often GPU- or streaming-flavored | Clean code under a performance framing, not LeetCode trivia |
| Domain deep-dive | Your strongest area: inference kernels, video models, or distributed systems | Depth over breadth; can you defend real design choices you made |
| Applied / take-home or live build | Optimize or debug something under a latency constraint | Profiling instinct, willingness to measure before optimizing |
| Founder / research-lead round | Vision, taste, why Decart specifically | Judgment, ownership, and whether you’d thrive in a small fast lab |
One logistics point for international candidates: Decart has not published a visa-sponsorship policy, and its footprint is split between Israel (Tel Aviv and a northern site) and the US (San Francisco and New York). Research and inference-systems work appears in both geographies, but the exact mapping of role bucket to office isn’t public, so confirm sponsorship, location, and remote flexibility with the recruiter on the first call rather than inferring it from the job post.
Questions phrased the way this domain asks them
No Decart loop has been posted publicly, so the shapes below are inferred from the domain, not candidate reports. A frame-budget question in some form is near-universal at real-time-inference shops, so plan on one. Something like: “You have 33 milliseconds per frame at 30 fps. Your model’s forward pass takes 45. Walk me through every lever you’d pull.” Good answers move fast across distillation to fewer steps, quantization, kernel fusion, caching what’s stable between frames, and cutting resolution or width, naming the quality tradeoff each costs. Weak answers just say “use a smaller model” and stop.
Autoregressive-drift questions are the other near-certainty anywhere real-time video is the product, since it’s the failure mode that defines the medium. “Your model generates each frame conditioned on the last. Over a few seconds the image degrades or hallucinates. Why, and how do you measure and slow it down?” They want error accumulation, teacher-forcing versus rollout during training, context windows over past frames, and concrete drift metrics rather than a hand-wave about “consistency.”
On the systems side, questions get specific about hardware: whether an operation is memory-bandwidth-bound or compute-bound and how you’d tell, how you’d structure attention or KV-cache for a stream of video tokens instead of a fixed prompt, and how you’d keep requests batched without any one missing its deadline. If you’ve written CUDA, expect to defend a kernel you optimized, including its memory-access pattern and occupancy math.
Two are worth seeing as strong versus weak answers, since they sit where inference and video meet. The first is the profiling prompt: “The GPU looks only 60% busy but frames still drop. Where’s the time going?” A strong answer treats the 60% as a symptom and hunts for the stall, host-to-device copies, kernel launch overhead, synchronization points, a preprocessing step stuck on the CPU, and asks to see a trace before guessing. A weak answer takes the number at face value and reaches for a bigger batch or a beefier GPU, which often makes tail latency worse.
The second is the serving-path design: “Build the serving path for a live video-transformation model with thousands of concurrent streams.” A strong answer separates the per-stream real-time path from batchable work, reasons about how much cross-request batching fits before someone misses a frame deadline, and plans for backpressure and graceful degradation (a cheaper model or lower resolution before dropping the stream). A weak answer designs a throughput-maximizing batch server as if it were an offline job and never names the deadline that makes this hard.
A few more to prepare for:
- Given a diffusion video model at 8 sampling steps, how do you get to 1 or 2 steps for live use, and what breaks?
- Port an inference kernel from one accelerator to another with different memory bandwidth; what do you re-tune?
- How would you evaluate whether Decart’s engine actually beats a general-purpose one on a specific workload?
The behavioral and founder rounds are lighter on process and heavier on fit. A small lab that ships fast wants people who own problems without a spec handed to them. Have crisp stories about a time you owned a hard performance problem end to end. A structured STAR answer works fine, but keep it concrete and technical rather than corporate.
Compensation, without invented numbers
There’s no reliable public comp data for Decart yet; levels.fyi has no submitted data points yet, normal for a company this new. Anchoring to the peer set this guide already links, venture-backed inference labs like Modal, Fireworks, and Together, senior ML-systems and inference engineers in the US market as of late 2026 generally land in the rough range of $180K to $300K base, with meaningful equity on top and a wide spread by seniority and location. Treat that as a peer-anchored estimate, not a Decart number. Israeli-market cash bases typically run lower than US figures, with equity making up more of the gap; the SF and NY offices track US norms more closely. Run your own numbers with a total comp calculator and read up on salary negotiation before quoting a figure to a recruiter.
The equity here is growth-stage, not an early-stage lottery ticket, and that changes which questions matter. At a reported ~$3.1B valuation your strike is set against a high last-preferred price, so the gap between strike and paper value is thin and the near-term upside rides on the next markup, not a 100x. Ask for the preference stack: several rounds raised in under two years means layered liquidation preferences sitting ahead of common in a modest exit. Ask about your dilution across those rounds and any planned raise, and get the strike, the last preferred price, and your share count in writing. When there’s no public comp data, the only real lever you have is a competing offer and a clear read on the level you’re being considered for.
One caveat on the company: funding details move fast and get garbled. The $100M round at a reported ~$3.1B valuation was announced in August 2025; some job aggregators list a larger 2026 round I couldn’t confirm in press, and trade reports have floated acquisition interest with nothing confirmed. None of it changes how you’d prep, but the numbers are a moving target if a recruiter cites them.
How to prep in the two weeks before
Rebuild your latency intuition first. If you can’t estimate the milliseconds a transformer forward pass takes on a given GPU, or explain the arithmetic-intensity roofline in a sentence, start there. Then get your hands dirty: pull an open-weights video model you can actually run, LTX-Video or CogVideoX are reasonable choices, and profile it with Nsight Systems or the PyTorch profiler until reading a trace and finding the stall is a reflex rather than a research project.
On the reading side, the distillation questions get easier if you know the lineage of few-step sampling: progressive distillation (Salimans and Ho, ICLR 2022), consistency models (Song et al., 2023) and the latent-consistency follow-ups, adversarial diffusion distillation (the SDXL-Turbo technique, 2023), and distribution matching distillation (DMD, 2024). For the drift side, diffusion forcing (NeurIPS 2024) and the recent causal video-distillation work aimed at streaming generation cover the exact regime Decart operates in. You don’t need to reproduce any of it; you need to say what each buys you and what it costs in quality.
For the broader landscape of how these labs run their loops, the AI-native company interview guides hub is the fastest way to calibrate expectations, and the AI-startup interview difficulty index gives a general sense of where real-time-inference roles land relative to other AI shops, though it doesn’t rank Decart specifically. The thing that separates a strong Decart candidate from a merely qualified one isn’t knowing more models. It’s the instinct, when the frame is late, to open a profiler instead of a bigger checkpoint.
Practice the behavioral round:
