Senior Research Engineer (Agentic Behavior)

At JetBrains, code is our passion. Ever since we started, back in 2000, we've been striving to make the strongest, most effective developer tools on earth. Today, AI-powered coding agents are becoming a core part of how developers write Kotlin – and we want to make sure they write it well.

The Kotlin AI Value Stream team is responsible for how AI agents understand, generate, and improve Kotlin code across all platforms: Android, Kotlin Multiplatform, server-side, web, desktop, and others. We build the evaluation infrastructure, error analysis tools, and post-training pipelines that measure and improve agent behavior on real Kotlin developer tasks. As a Research Engineer on this team, you'll own the end-to-end loop: Analyze how agents fail on Kotlin → build evals that capture those failures → research and implement methods to fix them → measure the improvement.

Your work will directly shape how millions of developers experience Kotlin through AI coding agents. As part of our team, you will: Build tools for agentic error analysis Design and implement tooling to systematically capture, classify, and analyse errors that AI coding agents make when generating Kotlin code. Build observability pipelines over agentic traces – mining patterns from agent sessions in JetBrains IDEs, Junie, Claude Code, Cursor, and other coding agents.

Build evaluation pipelines Design, implement, and maintain evaluation pipelines that measure Kotlin code generation quality across dimensions, including correctness, idiomaticity, build success, framework usage, and test coverage. Build simulation environments where coding agents can be measured on realistic Kotlin developer tasks – from greenfield KMP projects and Gradle dependency management to migrating Spring applications from Java to Kotlin. Own evaluation infrastructure: metrics, experiment tracking, automated regression checks, and reproducible benchmarking.

Research methods for improving agent and model behavior on Kotlin Experiment with post-training techniques (SFT, DPO, GRPO) to improve how models handle Kotlin-specific patterns, idioms, and frameworks. md files, compiler-as-verifier feedback loops, Kotlin LSP integration, and MCP-based tooling. Run experiments to measure impact: A/B comparisons, benchmark suites, and before/after analyses on real codebases.

Collaborate with model providers (Anthropic, OpenAI, and Google) to translate Kotlin-specific findings into model improvements. Build public Kotlin benchmarks Design and build open-source benchmarks that measure AI coding agent performance on Kotlin tasks and eventually become the standard reference for the ecosystem. Create task datasets covering the breadth of Kotlin usage: the server side (Spring, Ktor), multiplatform projects (KMP), build systems (Gradle), Android, library development, and others.

Include both mined real-world tasks and carefully designed synthetic tasks that test specific Kotlin capabilities. Maintain and evolve benchmarks as models improve, ensuring they remain challenging, relevant, and contamination-resistant. We'll be happy to have you on board if you have: Hands-on experience building evaluation or analysis pipelines for LLMs or AI coding agents in a research or production setting.

Strong Python engineering skills (at least three years), with the ability to write clean, maintainable code in data-heavy and ML-adjacent codebases. Experience with data analysis at scale: querying large datasets (SQL/Athena), building data pipelines, and performing statistical analysis of experimental results. The ability to own projects end to end – from identifying a problem in agent traces to designing an eval, running experiments, and shipping a fix.

A product-aware mindset: You care about how agents are actually used by developers and can translate real failure modes into evaluation and training work.