**About VinSmart Future**
*VinSmart Future (VSF) is the leading technology company within the Vingroup Corporation, formed by the merger of the group's entire technology ecosystem, including VinApp, VinIT, VinBigdata, and other tech units. As a core driver of Vingroup's future growth, VSF is at the forefront of technological development, with artificial intelligence (AI) as its foundation. With a talented team of nearly 4,000 local and international technology experts, VSF focuses on creating high-utility technologies that enhance lives and connect data, models, and infrastructure to unlock new possibilities.*
**AI Guardrail Model Lead**
***Work Location:***
- ***Ha Noi:***
*TechnoPark Tower, Vinhomes Ocean Park, Hanoi*
- ***Ho Chi Minh City:***
*Vincom Dong Khoi, HCMC*
**Job overview**
*We are looking for an experienced AI Guardrail Model Lead to lead the development of AI safety models that safeguard Large Language Models (LLMs) and AI agents across the Vingroup ecosystem. The role combines hands-on model development with technical leadership, covering model architecture, training, evaluation, and inference optimization. You will drive innovation in areas such as content safety, prompt injection detection, privacy protection, and policy-aware AI systems. The Lead is expected to bridge model research and production inference: defining model-service interfaces, streaming behavior and serving requirements together with backend/platform engineers, without owning the broader application or gateway platform. Working closely with platform, security, and product teams, you will ensure guardrail solutions are scalable, reliable, and production-ready. The role owns model behavior and the model-serving contract, and jointly designs inference integration with backend/platform teams. Platform teams own gateway/runtime execution, authorization, transport and production infrastructure; ownership of deterministic guardrail processing is defined per component. This position is ideal for a senior AI leader with deep expertise in LLMs, applied research, and mentoring AI/model engineers and collaborating closely with software/backend engineers.*
**Key Responsibilities:**
- **LLM model architecture**
. Set direction across decoder-only LLMs, discriminative heads and instruction-conditioned decision models. Compare specialized classifiers with shared multi-policy models using explicit quality/compute criteria
- **System-1 and dynamic policy reasoning**
. Research fast policy deciders, including Jev-style instruction-conditioned approaches. Enable policy bundles supplied at inference time, per-policy verdicts and generalization to unseen or updated instructions without retraining for every tenant
- **Streaming decision modeling**
. Lead causal and incremental classification; design prefix-level supervision, uncertainty handling, KV/prefix-cache reuse, state management and the balance between detection timeliness and latency
- **Inference systems and model-service architecture**
. Define how guardrail models are exposed as reliable low-latency services, including streaming and non-streaming inference, batching and cache reuse, state/session handling, model and policy versioning, error/overload behavior, and stable inference contracts. Work closely with backend and platform engineers on integration while keeping gateway/runtime implementation outside the core modeling role
- **Training methods and model improvement**
. Choose training objectives and recipes: supervised fine-tuning, LoRA/PEFT, multi-task learning, distillation, selective alignment/preference methods and calibration when technically justified
- **Safety and untrusted content**
. Guide research for unsafe content, jailbreaks, direct/indirect prompt injection, malicious tool/web input and sensitive-data detection; favor separate specialist components where they outperform a universal model
- **Evaluation and technical leadership**
. Own dataset quality, held-out policy tests, false-positive/failure analysis and controlled ablations. Review designs and code, mentor AI/model engineers and translate findings into clear model-service contracts
**Requirements**
**Must-have**
- Approximately 5+ years in ML/NLP/LLMs or compelling equivalent outcomes, including substantial first-hand model training and evaluation.
- Strong Python/PyTorch and Transformer/causal LM understanding: attention, tokenization, representation heads, training objectives, adaptation, decoding and inference behavior.
- Practical command of SFT, parameter-efficient fine-tuning, distillation and multi-task learning; ability to assess when more complex alignment or optimization techniques add value.
- Ability to design instruction/policy-conditioned models, structured per-policy predictions and experiments separating memorization from policy generalization
- Fluency in recall/FPR at fixed operating points, calibration, label quality, dataset leakage, distribution shift and reproducible ablations.
- Hands-on GPU profiling, batching and caching trade-offs; ability to reason about stateful streaming inference and production constraints.
- **Production inference understanding.**
Practical understanding of how ML/LLM models are served in production: synchronous and streaming APIs, batching, backpressure, cancellation, state/session lifecycle, model loading and versioning, observability, failure handling, and GPU-serving trade-offs. Able to design clear contracts and debug integration issues with backend/platform engineers; deep backend specialization is not required.
- Evidence of mentoring AI/model engineers, code/design review and cross-functional technical decision-making; able to use English research literature.
**Nice-to-have**
- Experience with low-latency System-1 decision models or Jev-style policy deciders, policy-as-instruction training, or streaming guardrail architectures.
- Work in model safety, prompt injection, privacy/PII, adversarial or multilingual data; Vietnamese NLP experience.
- Hands-on experience with production inference stacks such as vLLM, Triton, SGLang or equivalent; distributed/model-parallel serving, continuous batching, prefix caching, quantization or custom serving extensions.
- Research publications, open-source training contributions or strong reproducible modeling artifacts.
**Why You’ll Love Working Here:**
- Flexible working hours and attendance policy (Work from Home on working Saturdays).
- Attractive compensation and bonus packages, highly competitive in the market.
- Exclusive employee benefits across the Group’s ecosystem in accordance with company policies.
- Opportunity to work on large-scale and strategic technology projects.
- Professional technology environment with leading scientists, experts, and engineers from top technology companies in Vietnam and around the world.
- Free access to learning platforms such as Udemy, Coursera, and O’Reilly; internal workshops; sponsorship for professional certifications; and exclusive mentoring programs from the Group and Company leadership team.
- Full statutory insurance coverage in accordance with Vietnamese Labor Law (Social Insurance, Health Insurance, Unemployment Insurance), along with private healthcare insurance based on job grade and annual health check-ups at reputable hospitals and healthcare centers nationwide.
- Participation in internal activities, team-building programs, and annual company events.
