Causal Labs — Machine Learning Infrastructure Engineer

Causal Labs — Machine Learning Infrastructure Engineer Type: Full-time | On-site | San Francisco, CA Compensation: $200K–$400K + highly competitive early-stage equity Hiring count: 2–3 Visa sponsorship: Yes — open to visa transfers (e.g. OPT, H-1B transfers); can sponsor visas Reports to: Dar Mehta, Co-founder & CEO About Causal Labs Causal Labs is a research company building AI capable of (1) predicting the future and (2) identifying the optimal actions to change that future — what they call general causal intelligence. To get there, they are training a Large Physics foundation Model (LPM) to learn causality through physics, using weather as the initial proving ground because it is the most well-observed physical system with rapid, objective ground-truth feedback and data at a scale that dwarfs today's LLM training sets. The team comes from self-driving, drug discovery, and robotics. Founded: 2024 | Team size: ~10 (headcount conflict — see note) | Total funding: unannounced Series A (prior $6M seed led by Kindred Ventures, with Refactor, BoxGroup, and others) Industry: AI / physical AI / foundation models Website: www.causallabs.ai Office: South Park, San Francisco (5 min from Caltrain) Founders Kelsie Zhao (Co-founder & CTO) — self-driving veteran, built foundational components of Cruise's core self-driving stack; Stanford grad. Dar Mehta (Co-founder & CEO) — background across Google Research, Meta, Cruise, and a YC-backed robotics startup; Waterloo grad. Why Candidates Should Join Counterintuitive technical thesis: A bet that scaling LLM/video models to $1T of compute will not get us to super-intelligence. For candidates who want to bet on a non-consensus approach to intelligence. Best-funded Series A in physics models: Positioned as the single most well-funded Series A company building physics foundation models. A domain with real ground truth: Weather offers rapid, objective feedback and data volume beyond LLM-scale — a rare setting where infra work is measured against reality fast. Early, high-ownership seat: ~10 people scaling to a much larger research/eng org this year; infrastructure engineers own the training and inference backbone end to end. Intake Call Summary Company: Mission-driven toward general causal intelligence; building a large physics foundation model with weather as the training ground. Team: Founders Dar and Kelsie come from self-driving research (Cruise). ~8–10 today, plans to grow to ~35 by EOY, weighted to researchers and infrastructure engineers. Role need: ML engineers/researchers with large-scale foundation-model experience; backgrounds in physics, robotics, biology, or AI at the frontier. Candidate bar: Must have hands-on experience training large-scale foundation models and be mission-driven; multimodal data and large GPU infrastructure experience is essential. Work environment: In-office 5 days/week in SF; explicit emphasis on intensity and long hours (70–80h referenced in screening notes); fast-paced, research-focused. Comp & logistics: Relocation support provided; office sized for expansion to ~35. Process: 30-min cultural screen 1-hour technical full-day onsite. Ideal profile: Intentional and mission-oriented with a clear decision-making history; generalist mindset able to work across researchers and engineers. Pain points: Scaling the team fast without lowering the technical or mission-alignment bar. The Role Own the distributed training and inference backbone for a foundation model trained from scratch — clusters, pipelines, and low-level GPU performance across model scales. What You'll Be Doing Design, deploy, and maintain large distributed ML training and inference clusters. Build efficient, scalable end-to-end pipelines to manage petabyte-scale datasets and model training across the full ML lifecycle. Research and test training approaches including parallelization techniques and numerical-precision trade-offs across model scales. Analyze, profile, and debug low-level GPU operations to optimize performance. Stay current on research and bring new ideas into the work. Tech stack: Distributed training frameworks (FSDP, DeepSpeed), NVIDIA GPUs, control systems, Linux, FPGA, Python, C++, Kubernetes/Docker, GCP/AWS/Azure. Qualifications Seniority 2–10 years building large-scale ML infrastructure for core foundation models. [Required] Work Experience Experience building ML infrastructure for core foundation models (not fine-tuning) . [Must have] Worked at a science or physical AI company (e.g. self-driving, robotics, biology). [Required] (recently updated by client) Experience as a generalist, working across the ML lifecycle. [Strongly preferred] Hard Skills Deep expertise in optimizing large-scale training and inference workloads. [Must have] Proficiency with distributed training frameworks (e.g. FSDP, DeepSpeed). [Required] Low-level GPU performance optimization and debugging (CUDA + JAX). [Strongly preferred] (recently updated by client) Soft Skills Demonstrates high intentionality in career choices and mission-driven focus. [Required] Miscellaneous Must work in-person 5 days/week in San Francisco (relocation provided). [Must have] Traits to Avoid Experience is only in fine-tuning or deploying models for applications. Candidates looking for a slow-paced or 9-to-5 work culture. Short tenures (<1.5 years) at multiple recent companies without clear reasons. Role Details Salary: $200K–$400K Equity: Highly competitive early-stage equity On-site policy: In-person 5 days/week in the South Park, SF office (relocation provided) Visa sponsorship: Open to visa transfers (OPT, H-1B transfers); can sponsor visas Employment type: Full-time Location: San Francisco, CA Screening Questions Why are you interested in Causal Labs? (mission orientation) Are you actively interviewing with any other companies? Which ones, and on what timeline? How serious are you about joining a startup — judged against who you're actually talking to, not just stated comfort with 5 days onsite / 70–80h / startup intensity? Experience fit: walk through the core-model infra work directly (put specifics in notes). Can you be on-site in SF? If not, are you willing to relocate? What is your salary expectation (base)? How actively are you exploring new opportunities? Paraform Scorecard Nine Yes/No calibration items on the submission form (all binary — no "?" option is enabled for this role). Overall rating maps to the candidate's score against this JD: Excellent 85+, Good 75–84, Ok 60–74, Poor <60. Only one pending candidate may be marked Excellent at a time. Experience building ML infrastructure for core foundation models (not fine-tuning). [Must have] Deep expertise in optimizing large-scale training and inference workloads. [Must have] Must work in-person 5 days/week in San Francisco (relocation provided). [Must have] 2–10 years building large-scale ML infrastructure for core foundation models. [Required] Worked at a science or physical AI company (e.g. self-driving, robotics, biology). [Required] Proficiency with distributed training frameworks (e.g. FSDP, DeepSpeed). [Required] Demonstrates high intentionality in career choices and mission-driven focus. [Required] Experience as a generalist, working across the ML lifecycle. [Strongly preferred] Experience with low-level GPU performance optimization and debugging (CUDA + JAX). [Strongly preferred] Paraform Candidate Questions Submission-form questions requiring written answers. One-liner cap: 140 characters. HM explicitly prefers short, concise answers. Why do you think they'd be a great fit? (recruiter voice — David's perspective) Why are you interested in Causal Labs? (mission oriented) (candidate voice) Are they actively interviewing with any other companies? Which ones? Timeline? (candidate voice) Serious about joining a startup — judged on who they're actually talking to, not just stated comfort with 5 days onsite / 70–80h / startup culture. (candidate voice) Experience fit — explicitly ask the core-model infra questions and put answers in notes. (candidate voice) Base salary expectations. (optional) (candidate voice) Additional info fields: salary expectation (+ details if outside $200K–$400K), on-site / relocation confirmation, initial phone screen conducted (Y/N), source (optional), anything else (optional), attachments (resume required, not LinkedIn-generated). Interview Process Stage 1 — Submit candidate After submission, you're notified if the hiring manager wants to proceed. Stage 2 — Initial Call (30 min) Call with Justin or Dar. No technical content. Focused on intentionality, career narrative, mission orientation, and who else they're interviewing with. Stage 3 — Technical Screen Stage 4 — Onsite Day (1 day) 9am to ~7pm in the South Park office. Meet the whole team, lunch, a representative work problem, and close with the founders. Remote candidates get flights and lodging covered. Stage 5 — Offer Extended Stage 6 — Candidate Hired Ideal Companies & Backgrounds No discrete Ideal Companies list was present on the role page. Background signal below is derived from the intake call, ideal-candidate profiles, and rejected-candidate feedback. Physical / science AI Self-driving (e.g. Cruise, Waymo-adjacent), robotics, drug discovery / biology, physics-driven modeling. Large-scale ML systems / accelerators Foundation-model infra teams, ML systems / MLSys groups, AI accelerator companies (e.g. Cerebras, Etched, Annapurna Labs/AWS), large GPU-serving orgs. Ideal Candidate Profiles For reference only — do not source these specific profiles. Tong Wu — LinkedIn Staff Software Engineer @ ByteDance | GPU Serving, Optimization, Recommender Systems | Fremont, US Strong work on AI infrastructure Tech-lead experience but strongly IC-focused Kamil Khan — LinkedIn Supercomputing @ Etched | Ex-Cerebras, Ex-Google AI | SF Bay Area Strong track record shipping AI products in a startup environment Yuan (Ryan) Li — LinkedIn MLSys @ Annapurna Labs (AWS) | CMU / ZJU alum | Greater Seattle Area Top university Recent experience in a very similar domain with ownership across the product Rejected Candidate Feedback Explicit distributed-training infra: Only submit candidates who can clearly demonstrate hands-on FSDP/DeepSpeed experience and building large-scale foundation-model pre-training clusters (include GPU scale metrics). Core model focus: Candidates must be building core model infra, not just inference and post-training pipelines; check for proven low-level GPU profiling/optimization relevant to foundation models. Mission & domain fit: Prioritize physical-AI or science-driven backgrounds, and verify willingness for the 5-day onsite commitment and startup intensity. Observed rejection patterns (HM review, 41 rejections logged): the most common reasons are "lacks large-scale model training infrastructure experience," "focus not aligned with infrastructure role," "not a technical or skill fit," and "mission/company alignment mismatch." Most rejections happen at HM Review — screen hard on core-model infra and demonstrable FSDP/DeepSpeed + GPU-optimization evidence before submitting.