Benefits
- Health insurance
Skills
AI
Automation
Cloud Computing
C++
Data Pipelines
Developer Experience
Distributed Systems
Fine Tuning
LLM
Machine Learning
Nice to have
Full job description
Description
- We already operate a substantial agent system in production — a fleet of agents with an extensive library of skills and plugins, integrated into the tools our engineers use daily
- This role is about what it takes to make that system trustworthy, autonomous, and an order of magnitude more capable
- Three things sit at the center of it
- Closed-loop evaluation: Our ambition is to build the most rigorous evaluation system for AI work anywhere — closed-loop, meaning every agent action produces a measurable outcome that feeds back into whether that agent is trusted to act again
- Acceptance, revert, and override rates per workflow
- Statistical honesty about whether a difference is real
- Regression detection that fires before a human notices
- Everything else on this team depends on this being right, and almost nobody has built it well
- Agent platform: The runtime that makes autonomous agents safe to run against real systems: orchestration, sandboxing and isolation, tool and skill frameworks, memory, identity and permissioning, and the gateway and observability layer underneath
- Agents that touch production code and production infrastructure need containment and auditability before they need capability
- Autoresearch infrastructure: The automation of the research loop itself: agents that read the current state of a model and its metrics, form a hypothesis, launch an experiment, evaluate the result honestly, and either propose a change or discard the idea and move on
- At Nuro that loop runs against the training pipelines behind the driving model — real experiments, real compute budgets, real metrics that determine whether a behavior ships
- The hard parts are trusting the measurement, surviving experiments that take days, spending finite research compute wisely, and producing proposals a skeptical researcher can audit and reject
- Alongside this, the team builds agent-powered tooling across the engineering lifecycle — code generation, review, debugging, test and CI failure attribution, knowledge retrieval, triage
- There is also appetite on this team for post-training our own models where an internal workload justifies it, and the engineer in this role would be central to that work
- Build the closed-loop measurement layer that tells us, per workflow, whether agent output is accepted, reverted, or overridden — and use it to decide where autonomy expands and where it gets pulled back
- Take the autoresearch loop from assisted to unattended for a bounded class of experiments, including the eval and confidence machinery required to run it without a human in the loop
- Design the isolation and permissioning model that lets agents act on production repositories and infrastructure with an auditable record of what they did and why
Benefits
- Free Caltrain pass and commuter benefits
- Company stock options
- Work from home opportunities
- Health insurance
- Deep, current taste in LLM research. You understand how a model is trained from scratch — data, tokenization, architecture, pretraining dynamics, the full post-training stack of supervised fine-tuning, preference optimization, and RL — and you can reason about what a training decision does to model behavior. You follow the literature because you want to, not because it is on a roadmap
- You work end-to-end and do not need the problem pre-decomposed. This role has more surface than a specification
- Strong backend and distributed systems background at scale: cloud infrastructure, service design, storage, queuing, and the judgment to build things that stay up
- 5+ years of software engineering experience (or 4+ with a Master’s) in computer science, engineering, or equivalent practical experience. Staff-level candidates should bring correspondingly deeper scope and ownership
- You know what happens under the hood at inference. Attention and KV-cache behavior, batching and scheduling, quantization, speculative decoding, prefix caching, context handling, and how each trades off latency, throughput, and cost
- You have built and operated LLM-based agent systems in production — tool use, orchestration, sandboxing, retrieval, memory — and you know where they break
- Strong programming skills in Python
- You are opinionated about evaluation. You have argued with someone about whether a benchmark measured anything real, and you were right
- Hands-on post-training or fine-tuning experience — SFT, preference optimization, RL, distillation — including the evaluation work required to know whether it helped
- Familiarity with agent architecture patterns: planning, reflection, long-horizon memory, multi-agent coordination
- Go, C++, or Rust experience in addition to Python
- Experience with open tool-integration protocols, plugin or skill frameworks, and model-routing or gateway layers
- Background in developer experience or platform engineering, observability, or security isolation
- Experience with ML training or research infrastructure: experiment orchestration, evaluation pipelines, hyperparameter search, data pipelines
- Experience running inference serving, cost, or capacity at meaningful scale
Apply on company site
Opens Nuro’s own application page in a new tab.