Software Development Engineer (Machine Learning)

What you'll do

• Build and train guardrail models to detect prompt injections, jailbreaks, and sensitive data exposure.
• Design and tune a tiered detection cascade to balance accuracy against strict latency budgets.
• Fine-tune encoder and decoder models using distillation to move capabilities into deployable sizes.
• Optimize and serve models using quantization, compilation, and inference engines like Triton or vLLM.
• Conduct threat research on evasion techniques and convert new bypasses into training data and regression tests.
• Develop benchmark suites to measure detection rates against production-realistic false positive rates.
• Monitor deployed models for drift and maintain coverage for multilingual and regulated data.

What we're looking for

• Strong Python and production PyTorch experience; familiarity with Go, Rust, C, or C++ is valuable.
• Demonstrated experience training, fine-tuning, and evaluating transformer models using Hugging Face Transformers or equivalent.
• Production experience with inference serving systems like Triton, vLLM, TensorRT-LLM, or TGI including batching and memory tuning.
• Practical experience in model optimization techniques such as quantization, distillation, pruning, or graph compilation.
• Working knowledge of tokenization, text normalization, Unicode handling, and MLOps practices including Docker and Kubernetes.
• Experience in security or abuse-detection domains, LLM threat landscapes, or with OWASP Top 10 for LLM Applications (preferred).
• Experience with gradient-boosted trees, NER, PII detection, CUDA, or synthetic data generation (preferred).