Stochastic Modeling Intern

This is a paid internship for a graduate student or advanced undergraduate. You will work directly with Canyon Code's CEO on the data side of CanyonOS: characterising how multi-agent workflows actually run, from the runtime's own traces and from public datasets, and building the statistical models that turn those traces into predictions the platform can act on.

Agentic workflows are stochastic. Two runs of the same task can differ by an order of magnitude in tokens, latency and tool calls, and the people operating them have little more than averages to plan with. Your work is to make that behaviour measurable, modellable and predictable from small samples.

• Design and run experiments over agentic workflows: grids of tasks, prompts and configurations with repeated runs, on public benchmarks and on CanyonOS traces.
• Characterise run-to-run variability in tokens, latency and tool calls, and how it depends on inputs, prompts and configuration.
• Build probabilistic models of workflow behaviour, such as state-transition models and Bayesian hierarchical models estimated from small samples, and test their predictions on held-out runs.
• Write static analysis over Python agent code to extract its structure: agents, the calls between them, loops, retries and prompt templates.
• Reproduce and extend results from recent papers on agent token consumption, using public trajectory datasets.
• Work with the engineering team to bring what works into CanyonOS's telemetry and dashboard.

• Graduate student or advanced undergraduate in statistics, applied mathematics, computer science, electrical engineering or operations research.
• Solid probability and statistics: Markov chains, Bayesian estimation, hierarchical models, credible intervals, the bootstrap. You can derive a result, not only call a library.
• Have designed and analysed experiments: factorial grids, repeated runs, effects on a log scale, held-out evaluation.
• Fluent in Python with numpy, scipy and pandas, and comfortable working with large trace datasets.
• Have used a probabilistic programming tool such as PyMC, Stan or NumPyro.Plus
• Have built or closely used LLM agents, in a framework or in plain Python, and understand prompts, tool calls, context growth and token accounting, including cached versus fresh input.
• Can run agent benchmarks at scale: SWE-bench or similar harnesses, Docker, batch execution, cost tracking.Plus
• Can read other people's Python and reason about its control flow; experience with Python's ast module is a plus.
• Familiar with queueing theory or performance modelling.Plus
• Familiar with LLM serving stacks such as vLLM or SGLang, and with what prompt and KV caching mean for cost.Plus

Interested? Send your resume, a short note on what excites you about this role, and a link to something you have built or analysed.

Include Stochastic Modeling Intern in the subject line.