- Research and build Content Safety & Guardrails systems: input/output moderation, prompt injection & jailbreak detection, PII detection and redaction, topic control, hallucination mitigation.
- Build or fine-tune moderation/guard models (classifier, guard models like Llama Guard) optimized for Vietnamese and multilingual.
- Design guardrails framework integrated into agent/RAG pipeline and API serving (NeMo Guardrails, Guardrails AI or custom-built), optimizing safety layer latency.
- Build safety benchmarks and red-teaming processes: evaluation datasets, automated adversarial testing, monitoring new attack patterns.
- Balance safety and utility: measure and reduce over-refusal, fine-tune policy per product.
- Coordinate with product and legal teams to define content taxonomy and policy suitable for Vietnam and international markets.
Requirements
- 2 years (Middle) to 4+ years (Senior) NLP/ML Engineering experience; experience building text classification/moderation systems or working on LLM safety.
- Proficient in Python, PyTorch and transformers ecosystem; experience fine-tuning classification models and LLMs.
- Understand attack vectors against LLM systems: prompt injection, jailbreak, data exfiltration and corresponding defense techniques.
- Experience building datasets (labeling, synthetic data generation) and evaluating models with precision/recall trade-offs per product requirements.
- Experience deploying models to production: APIs, monitoring, continuous evaluation.
- Good English reading comprehension; understanding Vietnamese language and cultural context is a major advantage.
