RL Environments Engineer
RL Environments Engineer
Blockhouse, full time, NYC in person or hybrid preferred, remote within US time zones considered
About Blockhouse
Blockhouse gives institutional asset managers access to the kind of trading infrastructure that usually takes years and millions of dollars to build. Founded in 2022, we run a noncustodial platform where clients keep custody of their assets while they trade, hedge and compound portfolios across crypto and tokenized markets, from FX and equities to treasuries and compute futures. Alongside the platform we run live delta neutral funding rate and market making strategies with client capital across major venues, and we keep expanding into new frontier markets such as compute hedging.
The role
Frontier AI labs train and evaluate agents inside environments, which are realistic sandboxes paired with hard tasks and graders that decide whether an agent truly solved them. We're building these environments for quantitative research and trading systems engineering, drawing on the real work behind our live trading business. You'll own the environment as an engineer, designing the sandbox, the tasks and the graders, and making sure it all holds up when a lab's models try every shortcut they can find. This is an engineering role rather than a model training role, since the labs do the training, and we care far more about what you've built than about degrees or years of experience.
This isn't a speculative bet. We already work with frontier labs and have closed a high six figure AI training partnership built on our own in house research, and environments are our natural next step, so what you build goes straight into an active sales pipeline.
What you'll do
Design and build sandboxed environments for quant research and trading systems work, spanning data curation and feature work through backtesting and deployment
Write graders and reward signals that are hard to game, covering both exact checks and open ended outputs
Test environments adversarially, catching reward hacking, answer leakage and shortcuts before a lab does
Turn real research problems and engineering incidents from our team into well specified, long horizon tasks
Run frontier models through the environment with harnesses such as Harbor, and turn the results into benchmarks we share with labs
Own the reproducibility, packaging and sandboxing standards labs expect before they'll train on an environment
What we're looking for
You've built an RL environment, eval suite, benchmark or agent harness for LLMs, at a company, in research or as a serious side project, and can walk us through what made it hard
A working understanding of how RL training uses environments, including reward hacking, overfitting to graders and why task difficulty matters
Strong Python, since nearly every harness, grader and data pipeline here is written in it
Experience building Docker based sandboxes that restrict network access, cap resources and reset to the same state on every run
Comfort running agents at scale, using async code or job queues to launch hundreds of model runs in parallel and collect results cleanly
Hands on use of model APIs with tool calling, so agents can act inside an environment
Grader code that ranges from exact and tolerance based checks to rubric based LLM judging for open ended work
Rigorous testing habits, including fixed seeds, pinned data snapshots and tests that prove a grader passes right answers and fails wrong ones
Data handling with pandas or polars, Parquet and SQL across large datasets
Careful, skeptical judgment about whether a task is truly solved rather than whether a check passes
We hire on demonstrated work, so early career engineers with real environment experience are encouraged to apply
Nice to have
Enough Rust or C++ to read a real trading systems incident and turn it into a task
Familiarity with Harbor, Inspect or SWE bench style harnesses
Experience with GCP machine images and VMs
Experience in LLM post training, RLHF or reward modeling
A background in math, statistics or applied ML
Exposure to quantitative research, backtesting or market data, and curiosity about markets even without a trading background
Why Blockhouse
Ownership of a new product line from its first version
Close work with our quant researchers, who review task and grader design with you
Significant upside tied directly to the environments you build
Real, hard problems drawn from a live trading business
