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.

● Robust 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 ).

📌 Research Engineer (Data Synthesis) (Bengaluru)
🏢 Soch Street
📍 Bengaluru