The posting
THE ROLE
You'll turn difficult problems in real workflows into research questions, improve our models and the data and evals behind them, and carry results into production. You will own problems end-to-end with autonomy, treating evaluation and quantitative proof as first-class citizens in everything you create. This role is for a researcher with strong engineering ability who wants their work to impact real users.
We want well-rounded people, but most lean one way.
- Training & RL: You own the post-training recipe across SFT and RL on open-weight models. You've run training at scale and dealt with reward hacking and instability firsthand.
- Data & Evaluation: You own the data and evals models learn from and are judged by. Most of our tasks can't be checked automatically, so deciding what to measure, and spotting when a metric is wrong, is the research.
WHAT WE ARE LOOKING FOR
- Deep ML foundations: You understand learning, optimization, and statistics well enough to reason about model behavior and results.
- Hands-on experience: You've post-trained LLMs and owned the runs, or built data and eval systems close to model development. Ideally both.
- Research judgment: You turn ambiguous problems into clear hypotheses, design sound experiments, and can develop new approaches.
- Strong applied judgment: You stay close to real workflows and know when to improve the model, the data, or the system around it. You weigh quality, reliability, latency, and cost.
- Ownership: You help set direction and move from investigation, take responsibility for outcomes, and move from investigation to execution without waiting for a detailed plan.
- Clear communication: You communicate findings and uncertainty cleanly, seek context from domain experts, and help research, engineering, and product make decisions.
WHAT YOU'LL DO
- Own projects from problem definition through training, evaluation, and production.
- Pursue a research agenda to improve model capability on long-horizon reasoning and agentic workflows, including training and adapting models.
- Run SFT and RL experiments on leading open-weight models and debug training at scale.
- Design evaluations, benchmarks and verifiers that reflect real life sciences workflows, calibrate them against expert judgment, and investigate where they fail.
- Close the loop from product usage and customer feedback to new evaluations and training data, so models improve where customers need them most.
- Build the experimentation tools and pipelines needed to run, compare, and reproduce research efficiently.
- Partner with AI Strategy, product, and engineering to understand workflows and ship improvements to production.
NICE TO HAVE
- You have research publications, open-source contributions, or substantial industry work in ML, LLMs, agents, or evaluation.
- You have experience with reward modeling, synthetic data, or learning from human feedback.
- You have worked with distributed training, inference optimization, or large-scale experimentation.
- You have worked in a high-growth early-stage company.
Strong data and eval engineering counts even without prior LLM work, though LLM and RL experience is strongly preferred.
BENEFITS
- Competitive compensation, including meaningful equity.
- Medical, dental, and vision insurance for employees and dependents.
- Generous PTO policy.
- Paid parental leave.
We are hiring for this role across multiple levels.
At Valkai, we are committed to fostering a diverse and inclusive workplace. We provide equal employment opportunities to all employees and applicants without regard to race, color, religion, gender, sexual orientation, gender identity or expression, national origin, age, genetic information, disability, or veteran status.



