Research Engineer, Agentic Systems
ABOUT RAMP LABS
Ramp Labs is Ramp's AI research and engineering group. We post-train models on finance work, build the environments and evals they learn from, and ship them into the products 85,000+ companies use to move money.
ABOUT THE ROLE
Most agent work today is one agent, one trace, and a person reads the result. We're building systems where that stops working: thousands of agents fanned out over a large search space, each returning partial results, with an orchestrator that turns all of it into something useful: hundreds of merged PRs, a dataset, the input to another system, or something nobody has built yet.
You'll design and run those systems: how work gets split, how agents coordinate, how results get deduplicated, verified and ranked, and what it all costs. The problems are general (agent swarms, agentic map-reduce, long-running fan-out workflows) and you'll own them from prototype to production.
WHAT YOU’LL DO
- Decide how the work gets done: how to break a problem down, what each agent sees, how many agents to run, and which model handles which step
- Squeeze accuracy out of the models: context design, self-verification, voting and consensus across agents, and knowing when more agents stop helping
- Measure quality without reading every trace: sampling, automated verification, and precision and recall on the final output
- Drive down cost and latency per useful result: model routing, caching, early stopping, and smaller or fine-tuned models where they hold up
- Turn one-off experiments into reusable workflows on our existing orchestration stack
- Evaluate new models, inference setups and training approaches, and bring in what works
WHAT YOU NEED
- Experience building applied AI systems and LLM applications that shipped
- Hands-on experience with multi-agent or high-fan-out LLM workloads: parallel agents, map-reduce over models, large batch inference
- Strong engineering in Python; you build concurrent, distributed systems that hold up in production
- Judgment about output: you decide what the end result should look like before you scale the swarm, and you can separate signal from agent noise
- The habit of measuring: cost, yield, false-positive rate, and whether a change actually helped
- Ownership: you can run a roadmap and a production system with little oversight
- Clear written communication
NICE-TO-HAVE
- Built agent orchestration or job systems (Ray, Temporal, Modal or custom)
- Used LLMs to search large spaces: codebases, document sets, logs or other large corpora
- Post-training or hosting open-source models for high-volume inference
- Distributed systems background
- Public work we can look at: an agent framework, an open-source project, a write-up or a paper