Simile · San Francisco
OtherResearchFullTimeposted 5w ago
Simile is The Simulation Company. We simulate human behavior to keep people at the center of the decisions that shape the world. With AI, anyone can create a product, a campaign, a policy, or a script — the bottleneck has moved upstream. The hard question is no longer whether you can create something, but what to create, for whom, and how to bring it to life. Those are fundamentally human decisions, and they shouldn't be left to chance or handed off to an algorithm. We're building the infrastructure to understand human behavior at scale and to represent humans in an increasingly agentic world. Our mission is to simulate all eight billion people on earth.
We launched five months ago. Since then we've grown revenue 5x, built a new foundation model for human behavior that has run tens of millions of simulations for F100 enterprises, trained a first-of-its-kind confidence model that predicts the accuracy of every simulation, and released the first product that lets organizations verifiably predict the future. The world's leading companies use Simile to make business-critical decisions — from consumer leaders like CVS Health and Wealthfront to professional services organizations like Deloitte and Gallup — strategizing product launches, entering new markets, and forecasting earnings calls.
We've raised over $200M at a $2B post-money valuation led by Greenoaks, with Index Ventures, Hanabi, A*, Bain Capital Ventures, and CVS Health Ventures. We've grown from a small home in Palo Alto to a global team of 50+, and we're building a team of the best researchers, engineers, designers, and operators in the world. The future is too important to be left to chance.
Simile's Research organization builds the systems that every step of the model lifecycle runs on: data ingestion and schema design, distributed training, evaluation, serving, and monitoring. We are the reason a researcher's hypothesis can become a production simulation in days rather than quarters.
Two things make this problem unusual. First, our research-to-product pipeline is unusually tight - the experimental methods we validate on Monday are integrated into systems customers use to make high-stakes decisions. Second, simulating a society means running inference over populations of agents, not single requests. A single customer study can mean millions of model calls with interdependent state. Cost per simulation and latency per agent are not back-office metrics for us; they determine what research is even possible to run.
As a Member of Technical Staff in ML Systems, you will build the platform our researchers train, evaluate, and deploy on - and own it through the last mile, where a trained checkpoint becomes a production service serving millions of interdependent agent calls at a cost per simulation we can afford.
This is a role for someone who is energized by both halves of that. You will spend some weeks designing the data schemas and training pipelines a research team depends on, others profiling a serving path to find where the FLOPs and GPU memory are going, and others still bringing up cluster nodes or deleting the third redundant copy of a code path. The common thread is leverage: every improvement you make compounds across every researcher and every simulation we run.
We are looking for engineers who find it gratifying to see their work pushed to its absolute limits, and who own problems end-to-end - including the last mile of deployment that most people would rather hand off.
Build the ML platform our researchers live in. Design and operate the services, libraries, and tooling that cover the full lifecycle - data exploration, feature generation, experiment tracking, training orchestration, evaluation, and deployment. Success is defined by your ability to increase experiment velocity, streamlining the researcher’s path from ideation to a fully validated, production-ready model.
Make training and data pipelines fast. Own throughput end to end: model FLOPs utilization across our training configs, tokenization cost when the data mix changes, and ingestion paths that take hours today where they should take minutes. Profile where the time and GPU memory actually go, then fix it, including the observability that makes the next bottleneck obvious before it bites.
Make serving fast and cheap enough to run a society. Own the inference path our simulations run on: batching and scheduling, KV cache reuse across agents sharing context, quantization, and the request patterns unique to population-scale runs where one study is millions of interdependent calls. Cost per simulation and latency per agent decide what research we can afford to run at all, so treat them as research constraints, not ops metrics.
Scaling simulation Data. Lead the redesign of our data architecture to handle the complexity and sheer volume of our simulation models. You will define the schemas and ingestion logic to unify high-variety input streams (e.g., human survey behavior) and scale our training pipelines to meet the intense demands of society-scale modeling.
Own the GPU cluster. Keep a multi-node fleet healthy and saturated: node bring-up, topology-aware NCCL and RDMA configuration, scheduling and queue depth, storage lifecycle and checkpoint capacity, autoscaling of serving capacity alongside training jobs, and the alerting that tells us when GPUs are sitting idle. We run across more than one compute provider, so keeping the whole stack portable is part of the job.
Engineer scientific evaluations. Build evaluation tooling that goes beyond standard benchmarks, with rigorous statistical frameworks that prove the fidelity of our simulations - and that run fast enough to be part of the development loop rather than a gate at the end of it.
Push the state of the art. Reproduce, critique, and improve upon academic work in simulation, training, and inference optimization. Translate theoretical breakthroughs into production improvements, and document them with academic-level rigor.
Have deep systems and ML proficiency. High proficiency in Python and hands-on experience with modern ML frameworks (e.g., PyTorch, JAX). You can refactor a complex codebase for both performance and architectural integrity.
Understand modern ML architectures well enough to optimize them. You have an intuition for where the time and memory go and can act on it. You are comfortable around NVIDIA GPUs and the surrounding stack (NCCL, InfiniBand and NVLink topology, CUDA, Triton) or can get there quickly. Writing custom kernels is a bonus here, not the job.
Have built production ML platforms or MLOps systems. You have shipped the infrastructure other engineers and researchers build on: training orchestration, experiment tooling, model serving, or LLM application platforms. You know what makes the difference between a platform people adopt and one they work around.
Have architected, observed, and debugged production distributed systems. Bonus if they were performance-critical, and bonus again if you have had to substantially rebuild or refactor them as scale increased.
Own the deployment pipeline. You understand the training and fine-tuning lifecycle and can architect what continuous ingestion, monitoring, and high-availability serving actually require in practice.
Are research- and data-literate. You can navigate the ML research frontier, reproduce complex papers, tackle genuinely messy data states, and write about what you found with rigor.
Are self-directed and pragmatic. You figure out the most important problem to work on, pick up whatever knowledge you are missing to finish it, and know when to build the ideal solution versus when to adjust course.
Have a humble attitude, an eagerness to help your colleagues, and a desire to do whatever it takes for the team to succeed.
Have a strong quantitative foundation - typically a degree in Computer Science, Mathematics, Statistics, or a related field, though we care about demonstrated ability far more than credentials.
Experience optimizing inference for multi-agent or agentic environments, where requests are interdependent rather than independent.
Experience with distributed training at scale.
Experience building and shipping production AI agents, and familiarity with LLM serving frameworks (vLLM, SGLang, TensorRT-LLM) and inference-time optimization.
Interdisciplinary background in social science modeling or behavioral economics.
At Simile, we provide competitive compensation packages that include base salary, equity, and comprehensive benefits.
Salary Range: $200,000 – $400,000 USD
Note: Final offers are based on experience, specialized skills, interview performance, and relevant training.
Equity: Grants are available for eligible roles, subject to board approval.
Health & Wellness: Comprehensive medical, dental, and vision coverage.
Time Off: Flexible time off policies to support work-life balance.
We prioritize thoughtful conversations and clear examples of past work. Our hiring journey is designed to help both sides align on fit, working style, and expectations.
Reapplication Policy: To ensure a fair and thorough evaluation for all applicants, Simile observes a 90-day waiting period before reconsidering candidates for the same role.
Equal Opportunity: Simile is an equal opportunity workplace. We welcome applicants of all backgrounds and identities, valuing an environment where everyone can contribute authentically.
Accommodations: If you require support or reasonable accommodations during the application process due to a disability, please let us know. We are happy to assist.