Senior AI/ML Engineer (LLM / CV / OCR)

Dr. Kudos Lab Limited; Hong Kong; On-site; HKD 900,000 - 1,500,000; Full time

Job summary

Dr. Kudos Lab Limited in Hong Kong is building a foundational AI platform centered around locally or cloud-hosted LLMs and CV/OCR models.

The Senior AI/ML Engineer will own the full lifecycle from evaluation to deployment, integrating with MLOps and Backend teams to deliver production-ready AI foundations. The role is deeply technical, hands-on, and requires autonomy, strong coding depth, and cross-functional collaboration.

Qualifications

• Strong proficiency in Python and PyTorch.
• Hands-on experience with HuggingFace Transformers, PEFT, TRL, and related tooling.
• Practical experience fine-tuning LLMs (7B–70B parameters).
• Experience using Unsloth for efficient LLM fine-tuning and experimentation.
• Solid understanding of NLP fundamentals: tokenization, embeddings, attention mechanisms, model alignment.
• Experience with training optimization frameworks (DeepSpeed, Accelerate, FSDP).
• Familiarity with inference frameworks (vLLM, TGI, TensorRT-LLM).
• Experience with vector databases (FAISS, Milvus, Pinecone).
• Experience with CV/OCR frameworks (OpenCV, YOLO, PaddleOCR, Azure Vision, Google Vision).
• Strong engineering fundamentals: clean code, reproducibility, experiment tracking.
• Strong analytical and problem-solving ability.
• Ability to work independently and take ownership of complex tasks.
• Clear communication skills and the ability to explain technical concepts to non-technical stakeholders.
• Curiosity and passion for staying at the forefront of LLM and CV research.
• Comfortable working in a fast-moving, evolving technical environment.
• Proficiency in English is required. Cantonese or Mandarin is an advantage.

Responsibilities

• Evaluate open-source LLMs (Llama, Mistral, Qwen, DeepSeek) and recommend suitable architectures.
• Perform supervised fine-tuning (SFT), LoRA/QLoRA, and parameter-efficient training.
• Use Unsloth to accelerate and optimize LLM fine-tuning (QLoRA, 4-bit training).
• Build and maintain training pipelines with dataset prep, cleaning, augmentation and synthetic data.
• Develop evaluation frameworks to measure accuracy, hallucination rate, safety and performance.
• Optimize inference through quantization, batching, GPU memory tuning and model compression.
• Collaborate with MLOps to deploy models on local GPU servers or cloud.
• Implement embedding pipelines and retrieval-augmented generation components.
• Work with vector databases such as FAISS, Milvus, or Pinecone.
• Support document ingestion, chunking, indexing and retrieval workflows.
• Build LLM workflows using LangChain, LangGraph or similar frameworks.
• Design prompt templates, chains and multi-step reasoning flows.
• Develop and integrate CV/OCR models with production APIs and platforms.
• Partner with MLOps to ensure reproducibility, reliability and scalability.
• Stay up to date with the latest research and apply relevant advancements.

Skills

Python & PyTorch

Transformers

PEFT / TRL

LLM fine-tuning

Unsloth

NLP fundamentals

DeepSpeed / Accelerate

Inference frameworks

Vector databases

CV/OCR frameworks

English proficiency

Communication skills

Tools

OpenCV

YOLO

PaddleOCR

Azure Vision

Google Vision

Job description

We are building the foundational AI platform for our next-generation products, centered around a locally or cloud-hosted Large Language Model (LLM). As a Senior AI/ML Engineer, you will take ownership of the full lifecycle of LLM and CV/OCR model development — from evaluation and fine-tuning to optimization and deployment. This is a deeply technical, hands-on role. You will design, train, and optimize models, build supporting pipelines, and collaborate closely with MLOps and Backend engineers to deliver a production-ready AI foundation. The role does not involve people management, but requires strong technical depth, autonomy, and cross-functional collaboration.

LLM Development & Fine-Tuning

• Evaluate open-source LLMs (Llama, Mistral, Qwen, DeepSeek, etc.) and recommend suitable architectures.
• Perform supervised fine-tuning (SFT), LoRA/QLoRA, and parameter-efficient training.
• Use Unsloth to accelerate and optimize LLM fine-tuning (QLoRA, 4-bit training, memory-efficient workflows).
• Build and maintain training pipelines, including dataset preparation, cleaning, augmentation, and synthetic data generation.
• Develop evaluation frameworks to measure accuracy, hallucination rate, safety, and performance.

Inference Optimization & Deployment

• Optimize inference performance through quantization, batching, GPU memory tuning, and model compression.
• Work with inference frameworks such as vLLM, TGI, TensorRT-LLM, or FasterTransformer.
• Collaborate with MLOps to deploy models on local GPU servers or cloud environments.

Retrieval-Augmented Generation (RAG)

• Implement embedding pipelines and retrieval-augmented generation components.
• Work with vector databases such as FAISS, Milvus, or Pinecone.
• Support document ingestion, chunking, indexing, and retrieval workflows.

LLM Workflow Orchestration

• Build and maintain LLM workflows using LangChain, LangGraph, or similar frameworks.
• Design prompt templates, chains, and multi-step reasoning flows.

Computer Vision & OCR

• Develop and integrate CV/OCR models using OpenCV, YOLO, PaddleOCR, Azure Vision, or Google Vision.
• Combine CV/OCR outputs with LLM pipelines to support multimodal use cases.
• Build preprocessing, post-processing, and evaluation workflows for image-based tasks.
• Work closely with Backend engineers to integrate models into production APIs.
• Partner with MLOps to ensure reproducibility, reliability, and scalability.
• Participate in architecture discussions and contribute to the evolution of the AI platform.
• Stay up to date with the latest research and apply relevant advancements.Develop and integrate CV/OCR models using OpenCV, YOLO, PaddleOCR, Azure Vision, or Google Vision.

Requirements

• Strong proficiency in Python and PyTorch.
• Hands-on experience with HuggingFace Transformers, PEFT, TRL, and related tooling.
• Practical experience fine-tuning LLMs (7B–70B parameters).
• Experience using Unsloth for efficient LLM fine-tuning and experimentation.
• Solid understanding of NLP fundamentals: tokenization, embeddings, attention mechanisms, model alignment.
• Experience with training optimization frameworks (DeepSpeed, Accelerate, FSDP).
• Familiarity with inference frameworks (vLLM, TGI, TensorRT-LLM).
• Experience with vector databases (FAISS, Milvus, Pinecone).
• Experience with CV/OCR frameworks (OpenCV, YOLO, PaddleOCR, Azure Vision, Google Vision).
• Strong engineering fundamentals: clean code, reproducibility, experiment tracking.
• Strong analytical and problem-solving ability.
• Ability to work independently and take ownership of complex tasks.
• Clear communication skills and the ability to explain technical concepts to non-technical stakeholders.
• Curiosity and passion for staying at the forefront of LLM and CV research.
• Comfortable working in a fast-moving, evolving technical environment.
• Proficiency in English is required. Cantonese or Mandarin is an advantage.

Preferred Qualifications

• Experience with Unsloth for high-performance LLM training.
• Experience with LangChain, LangGraph, or LLM workflow orchestration.
• Experience deploying LLMs in production environments.
• Published models or datasets on HuggingFace.
• Experience with GPU optimization, quantization (GGUF, AWQ), or model distillation.
• Experience working in small, high-impact engineering teams.
• Exposure to MLOps tools (MLflow, Weights & Biases, Kubeflow).