To apply for this job, you need to complete both steps below:

STEP 1:

Please click the link to submit your application directly to the company: 

Careers

Your application will only be received by Recruiter if submitted via above link. 

STEP 2:

Kindly scroll to the bottom of this page and complete the short VinUni Tracking Form.

Filling out this form alone does not count as applying. Kindly remind this form is not part of the company’s application process. It only helps Careers, Alumni, Industry and Development (CAID) Department discover more opportunities and follow up in case of system issues.

Key Responsibilities:

• LLM fine-tuning and adaptation. Train and improve Transformer/decoder-only models using SFT, LoRA/PEFT, multi-task objectives, knowledge distillation and appropriate classification/structured-prediction heads.
• Dynamic multi-policy decision models. Implement instruction-conditioned safety deciders: inject policy instructions or policy bundles at inference time, return findings per policy, and test generalization to policies not seen during training.
• Streaming and stateful inference. Build incremental/prefix-level inference and evaluations; understand causal attention, KV-cache, policy-prefix reuse, context handling and full-response versus early-prefix decisions.
• Modeling alternatives. Prototype fast classification heads, instruction-following deciders and hybrid decision paths; compare cost, correctness, calibration and latency under the same test protocol.
• Data quality and failure analysis. Curate policy-specific and synthetic/adversarial examples; diagnose false positives, false negatives, ambiguous labels, cross-domain failures and train/test overlap.
• Safety detection tasks. Develop content-safety and prompt-injection detectors for user input and untrusted tool/web/retrieval content; support PII/secret and image moderation when applicable.
• Service-facing model engineering. Build and maintain model-facing inference adapters with stable schemas for streaming and non-streaming use, per-policy outputs, calibrated scores, status/version metadata and measured latency/throughput; work with software engineers to integrate them into guardrail services.

Requirements

Must-have

• Approximately 3+ years of applied ML/LLM engineering or strong equivalent evidence, with ownership of training and evaluation work. 
• Python, PyTorch, Hugging Face/Transformers, tokenization, self-attention, causal language modeling and modern fine-tuning. 
• Ability to implement/debug SFT and LoRA workflows, loss functions, classification heads, multi-task labels and reproducible training experiments. 
• Understanding of autoregressive incremental decoding, model state, caching, batching and profiling GPU memory/latency. 
• Recall/FPR, threshold tuning, label auditing, slice-based error analysis, held-out evaluation and controlled ablation studies. 
• Readable Python, tests, clean configs and ability to diagnose bugs in data preprocessing, model outputs and inference adapters. 
• Clear technical reporting and an ability to read English papers; works well with data, benchmark and platform engineers. 

Nice-to-have

• Experience with System-1 deciders or instruction/policy-conditioned guardrail models (including Jev-like approaches); prior Jev experience is not required. 
• Knowledge distillation, preference/alignment training, structured reasoning models or custom heads on decoder-only LLMs. 
• Streaming inference optimization, prefix caching, vLLM/Triton or production GPU-serving workflows. 
• Safety/red-teaming, multilingual Vietnamese data, privacy detection or multimodal moderation.