Research Engineering Manager, STEM
Turing · United States
Onsite · Lead / Principal · AI Research Scientist / Engineer · Posted 10 November 2026
Agile / ScrumAWSData qualityDjangoDockerFlask/FastAPIGCPKubernetesLLMs / GenAIMachine learningPostgreSQLPrototyping
About the job
About Turing
Turing’s mission is to accelerate superintelligence to drive real economic progress. Headquartered in San Francisco, Turing works with frontier AI labs to generate high-quality datasets, reinforcement learning environments, and frontier research benchmarks that improve model capabilities in software engineering, enterprise knowledge work, and advanced STEM reasoning. In software engineering, Turing is the largest and longest-running data provider in the category. Turing also works with Fortune 500 enterprises across financial services, life sciences, healthcare, retail, automotive, and CPG to build and deploy end-to-end agentic AI systems inside mission-critical workflows. By operating on both sides, Turing closes the loop between frontier research and enterprise deployment, turning real-world deployment signals into better data, evaluations, and more capable models. Learn more at www.turing.com.
The Role
Turing builds large-scale datasets and reinforcement learning (RL) environments that power post-training for the world's leading AI labs and enterprises.
The STEM Horizontal team produces the high-signal coding and STEM data frontier labs train on. We're looking for a Research Engineering Manager to lead the research engineering behind its next layer: synthetic data that measurably improves models, and new benchmarks that measure what today's evals miss.
You'll lead a team of research engineers, work side by side with research scientists, and stay close enough to the code to review an eval harness line by line or prototype a generator yourself when the team needs it. You'll also build the team: hiring research engineers is a core part of this role, not a side task. Ideally you've been on the other side, at a frontier lab, and know what makes a lab trust a dataset or a benchmark result.
- Synthetic data that earns its place. Generation, verification, and filtering pipelines where every sample is checked, and where we can show which properties of the data drive downstream gains, not just how much of it we made.
- Benchmarks that stay meaningful. New evaluations for capabilities current benchmarks miss, built with calibrated difficulty, contamination controls, and statistics strong enough that a lab can make a decision on the result.
Quality failures here are quiet. A subtle bug in a verifier doesn't crash anything; it ships thousands of wrong labels into someone's training run and costs weeks. If you want to build the team and the systems that catch that early, this is the seat.
What You'll Do
1. Build and Lead the Team
- Build the team: define the profiles you need, run hiring, and grow research engineers selected for research taste as much as engineering rigor.
- Run the team's operating rhythm: project selection, planning, experiment reviews, design docs, and clear ownership.
- Coach engineers on experimental design, code quality, and presenting results; give direct, regular feedback and run fair performance cycles.
- Define the standards the team works to: eval methodology, data quality bars, and reproducibility — every dataset and result can be regenerated from a commit, a config, and a seed.
2. Partner with Research Scientists
- Turn hypotheses into experiment plans together: what we're testing, the baseline, the ablations, and the result that would change our minds.
- Set the agenda with research scientists. You're expected to bring your own hypotheses, not only implement theirs.
- Take research prototypes to production scale without losing the scientific intent.
3. Build Synthetic Data Systems at Scale
- Own synthetic data pipelines that generate, verify, filter, deduplicate, and version data, combining programmatic generators, LLM-based generation, automated verifiers, and targeted expert review.
- Judge data by its effect on models: ablations and fine-tuning or RL experiments that show which attributes (difficulty, diversity, correctness, coverage) drive gains.
- Build quality systems that hold up at volume: gold sets, LLM-judge calibration against expert labels, sampling audits, and agreement metrics.
4. Create New Benchmarks and Evaluations
- Find where frontier models fail and where existing benchmarks have saturated, and design evaluations that measure what's missing.
- Own benchmark methodology end to end: task taxonomy, difficulty calibration, graders and verifiers, contamination controls, held-out splits, and confidence intervals.
- Ship eval harnesses, including agentic and sandboxed-execution evals, that are deterministic, reproducible, and trusted by internal teams and partner labs.
- Make the call on whether a result holds up: partner labs and internal teams act on your sign-off.
- Publish benchmarks, datasets, and findings with our research scientists at venues such as NeurIPS, ICML, and ICLR.
5. Stay Hands-On and Work with Frontier Labs
- Write code every week on the hardest, highest-leverage parts: core verifier logic, eval statistics, the pipeline component everything else depends on. Expect a third or more of your time on direct technical work.
- Be the person the team brings its hardest debugging and methodology questions to: why a judge disagrees with experts, why an eval number moved, whether a gain is real or noise.
- Act as a technical counterpart to researchers at frontier AI labs: scope requests, set acceptance criteria, and explain results and trade-offs plainly.
- Partner with Strategic Project Leads, delivery, and engineering teams so research advances land in client work and in the platforms the team builds.
Sample Projects
- Identify a capability gap, design and ship a benchmark for it, and see it adopted by partner labs.
- Stand up a synthetic data pipeline for long-horizon coding tasks with execution-based verifiers, and run the ablations that show which task properties drive RL gains.
- Calibrate an LLM-judge grading system against expert labels, map where it fails, and decide where humans stay in the loop.
- Design the experiment that tells a partner lab whether a two-point gain on their eval is real.
- Decide which of three competing synthetic data approaches the team should bet on, and design the experiments that settle it.
Tech Stack
- Research: Python · PyTorch or JAX · SFT · DPO · PPO / GRPO · RL with verifiable rewards
- Evals & data: LLM-as-judge · programmatic verifiers · sandboxed execution · gold sets · contamination checks
- Platform: Python services (Django / FastAPI) · PostgreSQL · Docker · Terraform · AWS or GCP
What We're Looking For
This job is posted on Turing's career site. Radarak links to the original posting: apply there.
📣 Get new jobs without checking back
RSS feed