Cartesia Interview Guide (2026): Voice AI, State Space Models, and Two Tracks

Cartesia is a San Francisco company building real-time, multimodal AI on state space models (SSMs), the architecture family behind S4 and Mamba. It was founded in 2023 by researchers out of Stanford, including Albert Gu, a co-inventor of Mamba, and Karan Goel, who serves as CEO, with roots in Chris Ré’s Hazy Research lab. Its flagship product is Sonic, a text-to-speech model built for speed, with reported latency under 90 milliseconds to first audio; the company has since shipped Sonic-3 and paired its voice models with Ink, a speech-to-text model, into a single real-time voice stack. Funding has climbed fast: a $27M seed led by Index Ventures, a $64M Series A in March 2025, and a $100M round in November 2025 backed by Kleiner Perkins, Index Ventures, Lightspeed, and NVIDIA, bringing total funding to roughly $191M.

Process

Public detail on Cartesia’s loop is thin, so treat this as a picture drawn from a couple of candidate reports and the company’s stage, not a fixed script. Expect a short cycle, often one to three weeks from first contact to a decision, which is faster than a big-company loop. Reports describe it opening with a conversation with a founder or the CEO, followed by technical rounds tuned to your track. Which track you land in shapes everything: research and research-engineering candidates should expect questions on sequence modeling and training, while systems and product candidates get hands-on building, and one candidate described a phone screen that involved implementing a React component in Replit. Across tracks, sharpen your coding so the timed rounds go smoothly, be ready for ML system design centered on serving and latency, and prepare a couple of behavioral stories about shipping under ambiguity.

What they actually ask

• State space models versus transformers: explain how SSMs like Mamba differ from attention-based models, and where linear-time, constant-memory recurrence helps for streaming audio and long context.
• Streaming and low-latency inference: how would you keep time-to-first-audio under 100 milliseconds for a text-to-speech model serving many concurrent sessions?
• Voice architecture: walk through how a modern TTS or speech-to-text system fits together, from tokenization to the audio decoder, and where quality trades off against latency.
• Distributed training: how do you train and debug a large sequence model across many GPUs, and what tends to break first as you scale?
• GPU and serving internals: reason about a CUDA or Triton kernel, or modify a serving engine like vLLM or TensorRT-LLM under load, for inference and systems roles.
• Strong general coding: a clean, timed implementation problem with clear reasoning about complexity, since every track expects real engineering.

Levels and comp (2026)

Cartesia does not publish salary bands, and levels.fyi currently has little populated data for it, so treat any number as a rough guide and check levels.fyi and recent Blind threads before you negotiate. Expect the usual well-funded startup mix: a competitive base that often sits below top FAANG cash, a meaningful equity grant, and standard health benefits. For senior engineers in San Francisco, cash bases at well-funded AI startups commonly land somewhere from the low-to-mid $100Ks up toward $200K or a bit higher, with equity as the real lever. Equity is where both the upside and the risk sit. After a $100M round, a grant is priced off a high valuation, so options can pay off well or come to little depending on the outcome. Ask about the share count, the current preferred price, the vesting schedule and cliff, and any early-exercise or extended post-termination exercise window before you sign.

Prep priorities

Start by using the product. Read Cartesia’s docs, generate some speech with Sonic, and pay attention to latency and voice quality, because that gives you concrete talking points and shows real interest. If you are aiming at a research or research-engineering role, get fluent in sequence modeling: understand S4 and Mamba, why linear-time recurrence matters, and how SSMs compare with transformers on long context and streaming. Be ready to discuss training large models in PyTorch, tensor programming, and inference optimization with CUDA or Triton. For systems roles, drill model serving: vLLM or SGLang internals, GPU memory hierarchies, batching, and profiling under load. For product and full-stack roles, keep your coding fast and practical and be ready to build a working feature end to end. Across every track, keep general coding sharp, but expect the bar to reward real engineering over puzzle trivia.

Frequently Asked Questions

Do I need to know state space models to get hired? For research and research-engineering roles, familiarity with SSMs like S4 and Mamba is a strong advantage and sometimes expected. For systems, infrastructure, and product roles, deep SSM knowledge matters less than strong engineering and a working grasp of how the models are trained and served.

Is this a LeetCode-heavy loop? Based on public signal, coding matters but the emphasis leans toward real engineering and domain reasoning rather than hard algorithm puzzles. Keep your fundamentals sharp and expect problems tied to what Cartesia actually builds.

How long does the process take? Expect a short cycle, often one to three weeks from first contact to a decision, which is quicker than a big-company loop. Specialized research roles can run a little longer.

What is the difference between the research and engineering tracks? The research track leans on sequence modeling, training, and generative models, and often wants an advanced degree or equivalent experience. The engineering and systems tracks lean on serving, inference optimization, and shipping production code. Both value strong engineering, so pick the track that matches your strengths and prep accordingly.