Website:
Job details:
What you'll own
● Design and build pipelines that generate, execute, and validate synthetic training data - code and specifications alike - using compilers, test harnesses, and formal verification tools, not LLM-judge scoring.
● Build and maintain reward models / verifiers for RL post-training (e.g. GRPO) on code and spec-generation tasks.
● Own data quality end to end: decontamination, filtering, coverage analysis, dataset documentation.
● Work closely with the post-training team to close the loop between data gaps and model failures.
● Contribute to internal tooling for large-scale synthetic data generation and evaluation.
What we're looking for
● Experience building synthetic or instruction-tuning datasets for code or structured technical documents (specs, requirements) - with a named, shipped dataset, tool, or paper you can point to.
● Direct experience with execution-based, compiler-based, or test-based data validation.
● Familiarity with RL post-training methods (e.g. GRPO, PPO, DPO) and/or SFT data pipelines.
● Strong software engineering fundamentals - comfortable in one or more systems languages (C, C++, Rust, or similar) and at least one scripting language.
● Based in or willing to relocate to Bangalore.
Nice to have
● Exposure to embedded, safety-critical, or hardware-adjacent software (automotive, aerospace, telecom, defense, or EDA tooling).
● Experience with formal verification tools (CBMC, Frama-C, KLEE, ESBMC) or compiler internals (LLVM, Clang).
Click on Apply to know more.