Research Engineer, RL Environments and Infrastructure

Hyphen Connect; San Francisco (CA); Hybrid; USD 180,000 - 240,000; Full time

Job summary

Hyphen Connect is seeking a hybrid Systems Engineer and AI Researcher to lead the design and implementation of our agent evaluation framework and post-training data pipelines.

You will architect sandboxed execution environments, high-throughput reinforcement learning feedback loops, and robust evaluation setups that handle non-deterministic agent outputs in enterprise settings. This role requires in-person collaboration at our SF office and a pragmatic problem-solving mindset.

Qualifications

• Production AI evaluations, RL loops, or sandboxed environments deployed in production.
• Proficiency in Python, Rust, C++, Go, TypeScript or CUDA; quick to pick up new tools.
• On-site in San Francisco required.
• Able to work with undocumented SDKs and distributed training setups.

Responsibilities

• Build stateful agent environments with 50+ step trajectories.
• Scale post-training and RL pipelines for dynamic policy optimization.
• Design enterprise benchmarks to detect drift, hallucinations, and loops.
• Optimize systems to minimize latency across distributed GPUs and sandboxes.

Skills

Production AI Experience

Systems Polyglot

Pragmatic Problem Solver

Tools

Python

Rust

C++

Go

TypeScript

CUDA

Job description

We are looking for a hybrid Systems Engineer and AI Researcher to lead the development of our agent evaluation framework and post-training data pipelines. You will design sandboxed execution environments, high-throughput reinforcement learning feedback loops, and robust evaluation setups that handle non-deterministic agent outputs.

Core Responsibilities

• Build Stateful Agent Environments: Design deterministically verifiable, stateful sandboxes (web, OS, API, database) where agents can execute 50+ step action trajectories safely.
• Scale Post-Training & RL Pipelines: Implement high-throughput post-training infrastructure (RLHF, Direct Preference Optimization, Process-Supervised Reward Models) for dynamic policy optimization.
• Design Enterprise Benchmarks: Formulate evaluation metrics and automated grading harnesses that catch agent drift, hallucination, and loops in realistic enterprise environments.
• Systems Optimization: Keep latency low and compute efficiency high across distributed GPUs and sandboxed runtime environments.

Requirements

• Production AI Experience: Track record of deploying evaluations, RL loops, or sandboxed agent environments into production.
• Systems Polyglot: Deep systems background (Python, Rust, C++, Go, TypeScript, CUDA)—you pick up new tools and frameworks within days.
• San Francisco On-Site: In-person collaboration at our SF office to iterate quickly with founders and domain experts.
• Pragmatic Problem Solver: Comfortable navigating raw paper implementations, undocumented SDKs, and custom distributed training setups.