Research Engineer — Reinforcement Learning

RESEARCH ENGINEER — REINFORCEMENT LEARNING You'll bring reinforcement learning to Firecrawl's core product — building the training infrastructure, reward pipelines, and fine-tuning systems that make our models meaningfully better at extracting, understanding, and structuring web data. This isn't theoretical RL research. You'll build your own training infra, run fast experiments, ship models to production, and bridge the gap between classical RL approaches and modern LLM agent systems. If you care as much about training throughput as you do about reward design, this is the role. Salary Range: $180,000–$290,000/year (Range shown is for U.S.-based employees. Compensation outside the U.S. is adjusted fairly based on your country's cost of living. You can explore how we calculate this here: https://www.firecrawl.dev/careers/compensation.) Equity Range: Up to 0.15% Location: San Francisco, CA or Remote (Americas, UTC-3 to UTC-10) Job Type: Full-Time Experience: 3+ years in applied RL, ML engineering, or model training — with production systems Visa: US Citizenship/Visa required for SF; N/A for Remote ABOUT FIRECRAWL Firecrawl is the easiest way to extract data from the web. Developers use us to reliably convert URLs into LLM-ready markdown or structured data with a single API call. In just a year, we've hit 8 figures in ARR and 100k+ GitHub stars by building the fastest way for developers to get LLM-ready data. We're a small, fast-moving, technical team building essential infrastructure superintelligence will use to gather data on the web. We ship fast and deep. WHAT YOU'LL DO Build training infrastructure and reward pipelines from scratch. Design and operate the systems that train and evaluate Firecrawl's models. You'll own the full loop — data collection, reward modeling, training runs, evaluation, and deployment. You build the infra yourself because you're the one who needs it to work. Fine-tune models to achieve state-of-the-art results. Take foundation models and make them dramatically better at web data extraction, content understanding, and structured output generation. You know how to get from "decent fine-tune" to "best-in-class" and you have the patience and rigor to close that gap. Bridge LLM agents and classical RL. The most interesting problems at Firecrawl sit at the intersection of modern LLM-based agents and classical RL techniques. You'll design reward signals for agent behaviors, apply RL methods to improve multi-step agent workflows, and figure out where traditional RL approaches outperform prompting — and vice versa. Run fast experiments and iterate. You design experiments that test meaningful hypotheses, run them quickly, and make decisions based on results. You don't spend weeks on experiment infrastructure before getting a single result. Speed of iteration is a core part of how you work. Communicate clearly to non-RL people. RL can be opaque. You translate your work into language that engineers, product people, and leadership can understand and act on. You know how to explain why a reward function matters without requiring everyone to read the paper. Collaborate closely with the team. Work directly with the Search/IR-focused Research Engineer and the engineering team to connect RL improvements with search, ranking, and the broader product roadmap. WHAT WE'RE LOOKING FOR Builds their own training infra and reward pipelines. You don't wait for an ML platform team to set things up. You build the training loops, reward models, data pipelines, and evaluation frameworks yourself — because you understand that infra choices directly affect the quality of results. You've operated GPU clusters, managed training runs, and debugged convergence issues in production. Can fine-tune models to SOTA. You've taken models from baseline to best-in-class on tasks that matter. You understand the full fine-tuning lifecycle — data curation, training dynamics, hyperparameter sensitivity, evaluation methodology — and you have the taste to know when a model is actually good versus when the eval is flattering. Bridges LLM agents and classical RL. You're fluent in both worlds. You understand PPO, RLHF, reward modeling, and policy optimization — and you understand how modern LLM agents work, where they fail, and how RL techniques make them better. You see connections between these domains that most people miss. Production-minded. You care about whether your models work in production, not just on benchmarks. You've deployed models that serve real traffic and made hard tradeoffs between model quality, latency, and cost. Research that doesn't ship isn't research that matters here. Runs fast experiments and communicates clearly. You'd rather run three rough experiments this week than one polished one next month. When you have results, anyone on the team can understand what they mean — no decoder ring required. Backgrounds that tend to do well: RL engineers at AI labs or applie