New opportunity

PhD Holder | AI Scientist (Computer Vision & VLM)

Reolink · Singapore

About this role

Reolink is seeking a PhD-level AI Scientist to join their team in Singapore, focusing on intelligent visual technology for security solutions. The role involves developing and optimizing machine learning and deep learning algorithms to address complex business challenges. Responsibilities include model training, evaluation, data preprocessing, and feature engineering. The successful candidate will collaborate with cross-functional teams to integrate models into business systems and drive product innovation. Research duties involve staying current with AI advancements, specifically in audio and video algorithms for surveillance, including target detection, feature extraction, tracking, and recognition. Documentation and reporting of experiments and workflows are also required. Requirements: PhD in Computer Science, Applied Mathematics, Electrical Engineering, Pattern Recognition, AI, Automatic Control, Operations Research, Biology, Physics, Neuroscience, or Statistics. Candidates must possess strong programming skills in Python and C/C++, proficiency in mathematical/statistical concepts, and hands-on experience with frameworks such as PyTorch, TensorFlow, or Caffe. Experience with models like CNN, RNN, LSTM, Transformer, and ViT is essential. Preferred qualifications include experience in neural network design, parameter tuning, model compression, and inference optimization on platforms including NPU, GPU, DSP, ARM, and mobile/server environments. The position is based near Tai Seng MRT, Singapore, and offers medical benefits, yearly bonuses, and performance bonuses.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

PythonCC++Machine LearningDeep LearningComputer VisionVLMCaffePyTorchTensorFlowMxNetCNNRNNLSTMTransformerViTNeural Network DesignModel CompressionInference OptimizationNPUGPUDSPARMTarget DetectionFeature ExtractionObject TrackingObject Recognition

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Algorithm Development for Surveillance
The core mandate is to research and implement computer vision algorithms specifically for security monitoring, including target detection, feature extraction, tracking, and recognition. You must demonstrate the ability to translate academic research into practical, high-accuracy surveillance solutions.
Model Optimization and Deployment
Beyond training, you are expected to optimize models for deployment across diverse hardware platforms, including NPU, GPU, DSP, and ARM. Prioritize demonstrating experience in model compression and inference optimization to ensure algorithms run efficiently on edge devices.
End-to-End Data Lifecycle Management
You will be responsible for the full data pipeline, from collection and cleaning to feature engineering and guiding data labelers. Success requires maintaining high-quality datasets that directly improve the performance and reliability of neural network models.

Likely challenges to think through

Balancing Research Innovation with Hardware Constraints
A significant challenge is implementing state-of-the-art models like Transformers or ViT while adhering to the strict computational limits of embedded security hardware. Prepare by discussing how you have previously navigated the trade-off between model complexity and real-time inference speed on resource-constrained platforms.
Cross-Functional Integration of AI Systems
Integrating complex AI models into existing business systems requires effective communication with non-technical stakeholders. Prepare to explain how you document experimental workflows and translate technical performance metrics into business value for product innovation.

Practice your answers

Technical practice

Optimizing Vision Transformers (ViT) for Edge Surveillance

What to coverDiscuss the trade-offs between model accuracy and computational latency when deploying ViT architectures on resource-constrained hardware like ARM or DSP. Suggestion: Focus on techniques such as knowledge distillation, pruning, or quantization to maintain performance while meeting real-time inference requirements for security cameras.

Handling Data Imbalance in Surveillance Target Detection

What to coverExplain your approach to feature engineering and data preprocessing when dealing with rare events or imbalanced datasets in video surveillance. Suggestion: Address how you would guide data labelers to ensure high-quality ground truth and what loss functions or sampling strategies you would employ to improve detection sensitivity.

Model Compression for NPU and DSP Deployment

What to coverDescribe your experience with model compression techniques such as weight quantization or structural pruning to fit deep learning models onto NPU or DSP platforms. Suggestion: Highlight your understanding of hardware-specific constraints and how you validate that the compressed model retains sufficient accuracy for production use.

Integrating VLM into Security Systems

What to coverHow would you leverage Vision-Language Models (VLM) to enhance the capabilities of traditional surveillance systems beyond simple object detection? Suggestion: Discuss the potential for natural language queries or semantic scene understanding and the challenges of integrating these large models into existing business systems.

Tracking and Recognition in Challenging Environments

What to coverWhat strategies do you use to maintain robust object tracking and recognition in scenarios with occlusions, varying lighting, or low-resolution video feeds? Suggestion: Discuss the use of temporal information (e.g., LSTM or RNN) and how you evaluate the stability of tracking algorithms under real-world conditions.

Inference Optimization on Heterogeneous Platforms

What to coverCompare the challenges of optimizing deep learning inference for server-side GPUs versus mobile-side NPUs. Suggestion: Focus on memory bandwidth, power consumption, and the specific software stacks or compilers required to bridge the gap between training frameworks like PyTorch and target hardware.

Neural Network Parameter Tuning

What to coverDescribe your methodology for systematic parameter tuning and hyperparameter optimization in deep learning models. Suggestion: Explain how you balance automated search methods with domain-specific intuition to achieve optimal convergence and performance.

Algorithm Documentation and Reproducibility

What to coverHow do you ensure that complex experimental workflows and algorithm developments are well-documented for cross-functional teams? Suggestion: Emphasize the importance of version control, clear reporting of evaluation metrics, and creating documentation that bridges the gap between research findings and engineering implementation.

Behavioural practice

Collaborating with Cross-Functional Teams

What to coverDescribe a situation where you had to explain a complex AI model or research finding to non-technical stakeholders to drive product innovation. Suggestion: Use the STAR method to highlight how you simplified technical concepts and secured buy-in for your proposed solution.

Tackling Complex Research Problems

What to coverProvide an example of a time you encountered a significant technical roadblock while developing an algorithm. Suggestion: Focus on your logical approach to diagnosing the issue, the research you conducted to find a solution, and the final outcome of your efforts.

Staying Current with AI Advancements

What to coverHow do you keep track of the latest SOTA implementations in computer vision, and how have you applied a recent research paper to a practical project? Suggestion: Describe your process for evaluating new technologies and the criteria you use to decide if they are suitable for implementation in a production environment.

Managing Data Labeling Workflows

What to coverTell me about a time you had to manage or guide a team of data labelers to improve the quality of your training data. Suggestion: Explain how you defined the labeling guidelines, monitored quality, and how the resulting data directly impacted the performance of your model.

Sharpen your knowledge

Topics to prepare

Vision Transformer (ViT) and Transformer Architectures
The role requires expertise in modern architectures for surveillance tasks. Review the mathematical foundations of self-attention mechanisms and their application to video data. Exercise: Implement a custom attention block in PyTorch and compare its computational complexity against standard CNN layers for a target detection task.
Model Compression and Inference Optimization
Reolink emphasizes deploying models on edge hardware like NPU, DSP, and ARM. Study techniques such as weight quantization, pruning, and knowledge distillation. Exercise: Take a pre-trained CNN model and apply post-training quantization to evaluate the trade-off between inference speed on a mobile-class processor and model accuracy.
Surveillance-Specific Computer Vision Algorithms
The position focuses on target detection, tracking, and recognition in video streams. Review state-of-the-art algorithms for real-time object tracking and feature extraction. Exercise: Design a pipeline for multi-object tracking that maintains identity consistency across frames in a high-occlusion environment.

Skills to revise

C++ for High-Performance Computing
Beyond Python, the role requires C++ for deploying models into production systems. Exercise: Write a C++ wrapper for a deep learning model inference engine to handle real-time video frame processing, focusing on memory management and minimizing latency.
Neural Network Parameter Tuning
The role involves optimizing models for specific business challenges. Exercise: Conduct a systematic hyperparameter search (e.g., learning rate schedules, batch size, or optimizer selection) for a detection model and document the impact on convergence speed and final mAP (mean Average Precision).

Your preparation checklist

Prepare a technical portfolio of surveillance-focused models
Curate a presentation of your past research or projects specifically involving target detection, object tracking, and feature extraction. Be prepared to explain how you adapted architectures like CNNs, Transformers, or ViTs to solve specific computer vision challenges in real-world surveillance scenarios.
Demonstrate model optimization and deployment expertise
Prepare case studies detailing your experience with model compression and inference optimization. Focus on your ability to deploy models across constrained hardware environments such as NPU, DSP, or ARM platforms, as this is a key differentiator for the role.
Review your proficiency in C++ and Python for production
Be ready to discuss your approach to writing production-grade code. Prepare examples of how you have transitioned research-level code (e.g., PyTorch/TensorFlow prototypes) into efficient, deployable C++ implementations for business systems.
Formulate questions on Reolink's hardware-software integration
Prepare insightful questions regarding the current challenges the team faces when integrating VLM or advanced vision algorithms into their existing security hardware. Ask about their pipeline for data labeling and how they balance model accuracy with real-time inference requirements.
Synthesize your research experience with business impact
Draft a summary of how your PhD research can be applied to Reolink's product roadmap. Connect your theoretical background in areas like pattern recognition or statistics to the practical goal of enhancing intelligent visual technology for consumer security.

Availability last checked: 18 Oct 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…