About the Role
Cake by VPBank is one of Vietnam's fastest-growing digital banks — 6M+ users and on a mission to become the NextGen AI Bank. We believe generative AI will redefine banking: smarter, faster, and more human.
As a Senior GenAI Engineer, you'll drive our generative AI capabilities — LLMs, foundation models, and multimodal systems — across the full lifecycle: from adapting frontier models to shipping reliable, low-latency AI services for millions of users. Expect genuinely hard, Vietnamese-first problems, your own physical GPUs, and the freedom to bring the latest techniques into production fast.
Key Responsibilities
- Build or adapt models for Cake's own use cases — not just call third-party LLM APIs.
- Fine-tune and align LLMs and foundation models for real banking use — conversational and agentic assistants, fraud & risk, document intelligence, and Vietnamese-language understanding.
- Build end-to-end pipelines: data preparation, fine-tuning and alignment (SFT; PEFT — LoRA/QLoRA; preference optimization — DPO/SimPO/KTO; RL with verifiable rewards — GRPO/RLVR; agentic / tool-use RL — executable environments; distillation; model merging), rigorous evaluation, and deployment at scale.
- Optimize models for latency, cost, and reliability in production — serving, quantization, guardrails — and monitor them over time.
- Explore new architectures and training approaches, and bring emerging advances into production quickly.
- Champion responsible AI: fairness, safety, and compliance with banking regulations.
- Collaborate across product, backend, data, and MLOps teams — and mentor junior engineers.
Qualifications
- Bachelor's or Master's in Computer Science, AI/ML, NLP, or a related field — or equivalent hands-on experience.
- 3+ years in AI/ML with hands-on generative AI — LLMs and/or other modalities such as speech/audio or multimodal.
- Fluency with modern deep-learning frameworks (PyTorch, Hugging Face).
- Solid grasp of transformer architectures and modern adaptation — SFT, parameter-efficient fine-tuning (LoRA/QLoRA), preference optimization (DPO/SimPO/KTO), RL with verifiable rewards (GRPO/RLVR), and RAG.
- A proven track record deploying models to production — not just experiments or notebooks.
- Clear understanding of how foundation models are trained end-to-end: data curation, tokenization, pre-training objectives, and scaling.
- Strong problem-solving, a product mindset, and excellent cross-functional collaboration.
Nice-to-Have
- Pre-training or large-scale fine-tuning of foundation models from scratch.
- Speech/audio or multimodal generative experience — speech-to-speech, TTS/ASR.
- Publications at top venues (NeurIPS, ICLR, ICML, ACL, CVPR…) — or strong open-source contributions, competition wins, or well-known technical projects.
- Production optimization: modern serving stacks (vLLM, TGI, SGLang), quantization, and latency/cost tuning.
- Cloud & containerized deployment — GCP preferred; AWS or Azure welcome; Kubernetes a plus.
- Vietnamese or low-resource NLP, or fintech / digital-banking experience.
Why You’ll Love Working at Cake
• Build generative AI systems used daily by 6+ million users — your work makes an immediate, tangible impact.
• End-to-end ownership: from foundation-model research to production deployment.
• Work with a cutting-edge cloud-native stack (GCP, Vertex AI, Kubernetes, Airflow) and large-scale GPU compute.
• Collaborative, high-performing tech culture where experimentation and innovation are encouraged.
• Competitive compensation and the opportunity to grow within a fast-scaling digital bank.
Our Benefits
• Competitive compensation including a 13th-month wage and up to 3 months of performance-based bonus.
• MacBooks are supplied to all technical team members.
• BE Corp budget (varies by level) for transportation, food, and car bookings in the Be application.
• Social insurance contribution amount based on individual level.
• Annual health checks and premium medical healthcare (PTI) after probation.
• 15 days of annual leave for all staff.
• Company trips, team-building activities, and happy-hour events on a quarterly or annual basis.
