We're looking for a Principal Applied Scientist to lead the design and build of an evaluation framework for an enterprise planner-based agent. You will work directly with the client's engineering team to benchmark their agent across planner models, executor LLMs and inference optimization levers. Every proposed change, whether a model upgrade, a harness change or a routing strategy, becomes a controlled, statistically sound experiment with a measurable dollar outcome.

This is a research-heavy, customer-facing role with a significant build component. You will own both the experimental method and the tooling that makes it repeatable: a scenario-driven user simulator, isolated experiment execution, statistical comparison, and automated insights in a locally hosted dashboard the client can adopt as their own. The work runs in three overlapping streams (evaluation pipeline, grading and statistics, dashboard and experiment campaign) and ends in a recommendation the client can act on.

Why This Role Matters

At Robots & Pencils, we design AI systems for a human world. Our name says it all. Robots and pencils means engineering paired with creativity, because every agent we ship has to work for real people in real workflows. That balance is baked into how we operate.

Agents are only as good as the evidence behind them. Today, most teams pick a model or switch on a cost lever on instinct. Here, you replace instinct with data: you build the system that tells the client which planner model, which executor and which levers to roll out, how much quality each change costs or saves, and what it is worth in dollars at production volume. You also show how well the planner carries over from data analytics to tasks. This is a research role with real build accountability, and your work becomes the client's own decision-making tool. It demands people who take ownership, lead with craft, and care deeply about what they put their name on.

What You'll Do

Evaluation Pipeline & Simulation

Design the end-to-end evaluation workflow for a planner-executor agent: scenarios, user simulation, configuration, execution, grading, comparison and insights
Define the scenario schema, grading spec and validators with the client's team, who own scenario creation, so their data plugs directly into the framework (each scenario has a goal, persona, starting state and expected end state)
Build an LLM-based user simulator with a hidden goal and a stop rule that holds realistic multi-turn conversations, and check its realism against individual production logs
Build isolated, parallel experiment execution: config files, one run per config and scenario, and a harness adapter that routes all model traffic through the client's existing LLM gateway
Package optimization levers as plug-ins and implement a shortlist of 3 to 4 for v1, such as cascade-on-failure, cache-aware model switching and prompt caching
Tag every model call at the gateway and send traces to a trace store, so tokens, cache usage, latency and cost are attributed to each request

Grading, Statistics & Insights

Grade each run against the expected end state as the primary signal, with optional LLM judges for answer and plan quality
Compute the core metrics: verified success, cost per success, p95 latency, and failures traced to pipeline stage
Compare experiments on the same scenarios with paired methods: confidence intervals, win/loss/tie per scenario, and a non-inferiority test that a cost lever does not drop quality past an agreed margin. Say clearly when a result is inconclusive
Convert differences into projected dollar savings at production volume using real deployment costs, not just list token prices, and plot the cost-versus-quality frontier
Build an automated insight engine where a statistics layer computes every number and an LLM only writes them up, so each ranked finding links back to its evidence

Dashboard, Experiments & Readout

Build the locally hosted dashboard (Launch, Live, Results, Compare, Insights and Traces views), API-first with one-command local deployment so it can be integrated into the client's platform later