About the role
About Tonic Tonic builds the data infrastructure behind modern AI. We generate the synthetic environments that agents are trained and tested in, and we de-identify real enterprise data so it can be used safely in training and evaluation. Eight years in, we work with frontier AI labs pushing the edge of what models can do, and with hundreds of enterprises including Fidelity, Comcast, eBay, and Vanguard, on the data problems that sit at the center of where AI is going next.
About
The Role The models you build here are load-bearing. The environments you generate decide whether an agent is ready to ship or only looked good in a demo. The synthesis and de-identification models you train decide whether a bank can safely put its data near a model at all. And the work spans real range: in one week you might build evaluation that separates the best models from the rest on real tasks, train a synthesis model where both fidelity and downstream utility have to hold, and improve entity detection on messy production data. Real enterprise data, real stakes, and problems that don’t have textbook answers yet.
What You'll Do Design and build the systems that generate longitudinally coherent synthetic environments for agent training and evaluation, including persona modeling, task generators, and verifiable ground truth. Build and maintain synthesis models that generate realistic replacement values at very large scale, preserving format, statistical distribution, and semantic consistency so de-identified data stays useful downstream. Train and improve the NER models behind our entity detection, driving accuracy and recall across free text, structured fields, and mixed enterprise data at scale. Build evaluation infrastructure that grades agent outcomes, not just traces, and produces real discrimination between frontier models on real tasks. Fine-tune and evaluate open-weight models on Tonic-generated data, and turn benchmark results into product and research direction. Expand coverage into new domains, languages, and entity types, and handle the long tail of formats and edge cases that real customer data throws off. Own model evaluation across the board: precision and recall on detection, utility preservation on synthesis, and outcome-level grading for agents. Optimize inference so models run efficiently on large volumes of sensitive data inside customer environments. Partner directly with frontier labs and enterprise ML team to turn hard data problems into shipped model improvements. Set technical direction for a small, senior team and raise the bar on rigor, reproducibility, and shipping. What You’ll Bring 8+ years (or PhD with 3+ years) building production ML systems, with real depth in some combination of LLMs, agents, RL, NER, or information extraction. Hands-on experience training and shipping models to production, and a pragmatic bar for quality: you know how to measure it, where it breaks, and when it's good enough to ship. Experience with generative or synthesis models where output fidelity and downstream utility both matter, not just plausibility. Strong software engineering fundamentals. You write code others build on. Fluency with modern training and eval stacks (PyTorch, distributed training, standard agent and benchmark frameworks). Comfort working with messy, sensitive, real-world data and the privacy constraints that come with it. A track record of framing ambiguous problems and driving them to measurable, shipped results.