Member of Technical Staff - Pre-Training Evals Infra

Reflection is a research lab making intelligence open and accessible for everyone to use, customize, and build on. We build open models that let anyone control their intelligence and help shape the future of AI. Our mission: make intelligence open and accessible to all.

You will build the evaluation infrastructure that researchers use to understand model capabilities, guide training decisions, and assess readiness for release. The role spans evaluation authoring, distributed execution, and results analysis, with a particular focus on pre-training. You will work directly with researchers to make evaluations faster, more reliable, and easier to use, ensuring that the results accurately reflect the models being evaluated.

• Design and evolve evaluation abstractions. Build clear APIs and data models for evaluation tasks, datasets, configurations, model outputs, and scores, making it straightforward to add new evaluations and extend existing workflows.

• Build scalable, reliable execution infrastructure. Develop distributed systems for evaluating model checkpoints throughout training and ahead of release, with robust orchestration, observability, and recovery from failures.

• Optimize evaluation and inference performance. Profile inference servers and distributed execution to improve throughput, GPU utilization, and time to results.

• Establish evaluation correctness and reproducibility. Build validation and regression tests, version datasets and configurations, and preserve the provenance needed to reproduce results and investigate discrepancies.

• Build researcher-facing tools. Create intuitive interfaces and dashboards for inspecting model outputs, comparing checkpoints and experiments, and debugging unexpected behavior.

• Operate and monitor evaluations as training progresses. Track evaluation workloads and results, triage failures, and develop alerts and diagnostics that help distinguish changes in model quality from inference or evaluation issues.

• Turn research needs into reusable systems. Partner directly with researchers to identify bottlenecks, prioritize high-impact improvements, and turn bespoke workflows into self-service tools that work across teams.
  

• Strong proficiency in Python, C++, or Rust, with the ability to reason deeply about concurrency, memory behavior, and performance.

• Experience designing, building, and operating reliable, maintainable distributed systems, with sound judgment about interfaces, failure modes, and architectural tradeoffs.

• Working knowledge of model inference and evaluation, including how implementation and configuration choices can affect measured model quality.

• Ability to investigate complex problems across system boundaries, follow evidence, and identify root causes rather than just address symptoms.

• Strong product judgment and experience translating the needs of researchers or other technical users into clear APIs and usable tools.

• Comfort owning ambiguous problems end to end, prioritizing competing research needs, and balancing rapid delivery with long-term reliability and maintainability.

• Attention to detail and the persistence to investigate subtle discrepancies in model outputs, evaluation results, and system behavior.