New opportunity

[2027 Campus Hire] AI Algorithm Engineer

MediaTek · Singapore

About this role

This Singapore-based campus-hire role develops and optimizes AI algorithms, including models for mobile and embedded devices. The posting lists no work experience requirement.

Responsibilities

Develop generative AI language and multimodal model algorithms, including transformer-based architectures such as LLaMa and RNN- and state-space-model-based architectures such as Mamba and Linear Attention.
Develop computer vision, voice and NLP algorithms, and optimize Edge AI algorithms for APU use.
Work on neural network optimization through quantization, pruning and low-bit techniques; neural network design automation; reinforcement learning; and on-device learning.

Requirements

Strong understanding of AI models, especially transformers and LLMs, including graph structure, operators and inference behavior.
Familiarity with on-device frameworks for mobile or embedded devices, including TFLite and ExecuTorch.
Python, PyTorch and C/C++ programming skills.

Preferred qualifications

Knowledge of deep learning and generative AI algorithm or model development, optimization and training methodology.
Experience developing deep learning and generative AI algorithms for mobile or embedded devices is an advantage.

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

Skills for this role

Generative AILarge Language ModelsMultimodal AITransformersLLaMaRNNState Space ModelsMambaLinear AttentionComputer VisionSpeech AINatural Language ProcessingEdge AINeural Network OptimizationQuantizationPruningLow-bit TechniquesNeural Network Design AutomationReinforcement LearningOn-device LearningTFLiteExecuTorchPythonPyTorchCC++

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

Generative model architecture
The role develops language and multimodal algorithms across transformer, RNN and state-space approaches. The posting specifically calls for understanding model graphs, operators and inference behavior.
Edge AI development
Developing and optimizing algorithms for APU use is a stated duty. Familiarity with mobile and embedded frameworks, including TFLite and ExecuTorch, is required.
Neural network efficiency
Quantization, pruning and low-bit techniques are explicit areas of work. Prepare to explain what each changes in a model and how its effect on inference might be evaluated.
Algorithms across modalities
The posting covers vision, voice and NLP alongside multimodal models. Prepare examples of how algorithm choices and evaluation differ across these tasks, without assuming equal emphasis on each.
Implementation and emerging methods
Python, PyTorch and C/C++ are required. Neural network design automation, reinforcement learning and on-device learning are also listed duties; prepare to discuss relevant fundamentals rather than presume a particular project.

Likely challenges to think through

Balancing model quality and device efficiency
Likely preparation challenge, not a confirmed MediaTek problem: quantization, pruning or low-bit methods may reduce resource use while changing model behavior. Prepare a concrete comparison plan covering accuracy, model size and inference performance.
Moving from model graphs to on-device execution
Likely preparation challenge: an algorithm that works in PyTorch may behave differently when represented through operators in TFLite or ExecuTorch. Prepare to trace an inference graph, identify an operator issue and describe how you would check the deployed result.
Choosing depth across a broad research scope
Likely preparation challenge: the posting spans several architectures, modalities and advanced methods, making breadth-versus-depth a practical trade-off. Prepare one well-explained algorithm example, then show how you would assess whether its approach transfers to another listed area.

Practice your answers

Technical practice

How would you trace an LLM’s computation graph during inference?

What to coverThe posting emphasizes transformer and LLM graph structure, operators and inference behaviour. Explain the main operators, how data moves through the graph, and where you would look for an inference bottleneck.

How would you compare transformer, RNN, Mamba and linear-attention approaches?

What to coverThese architectures are named in the posting. As suggested preparation, compare how they handle sequence context and discuss trade-offs that could matter for inference on a mobile or embedded device.

How would you develop and evaluate a multimodal generative-AI model?

What to coverThe role includes language and multimodal algorithm development. As suggested preparation, describe how you would represent and combine inputs, choose a training approach, and test whether each modality contributes useful information.

What changes when developing algorithms for vision, voice or NLP tasks?

What to coverThe posting names all three areas. Compare their input representations and evaluation needs, then explain which parts of a model-development workflow could be shared and which would be task-specific.

How would you optimize a model for Edge AI on an APU?

What to coverAPU algorithm development and optimization are stated duties. As suggested preparation, explain how you would measure accuracy, memory use and inference performance, then decide whether a proposed change improves the relevant trade-off.

How would you choose among quantization, pruning and low-bit techniques?

What to coverAll three techniques appear in the posting. Explain what each changes in a neural network, how you would measure its effect on accuracy and inference, and how you would investigate a large regression.

How would you investigate a model that behaves differently in PyTorch and an on-device framework?

What to coverThe requirements name PyTorch, TFLite and ExecuTorch. As suggested preparation, describe how you would isolate differences in operators, model conversion or numerical behaviour without assuming a particular deployment pipeline.

Where would Python, PyTorch and C/C++ fit in an algorithm-development workflow?

What to coverThese are the programming skills listed for the role. Explain how you might use them for model experimentation and implementation, and how you would check that a faster implementation preserves the intended outputs.

How would you evaluate neural network design automation?

What to coverNeural network design automation is a stated work area, but the posting does not specify a method. As suggested preparation, outline a search objective, a way to compare candidate designs fairly, and checks against selecting a design that only looks good in limited tests.

What challenges do reinforcement learning and on-device learning introduce?

What to coverThe posting identifies both as advanced algorithm areas. As suggested preparation, distinguish their objectives and discuss how you would evaluate learning outcomes while accounting for the constraints of an on-device setting.

Behavioural practice

Tell me about a time you investigated an unexpected model result.

What to coverUse a real course, research or project example. Describe the situation, the checks you performed, and what the result showed; this is suggested practice for the role’s emphasis on model and inference behaviour.

Describe a time you made a trade-off while improving an algorithm.

What to coverGive the situation, the alternatives you considered, your action and the measured result. As suggested practice for the optimization duties, be clear about what improved and what, if anything, got worse.

Tell me about a time you learned an unfamiliar model or framework to complete a task.

What to coverDescribe why you needed to learn it, how you tested your understanding, and the outcome. This is suggested practice given the range of architectures and on-device frameworks named in the posting.

Describe an experiment that did not support your initial approach.

What to coverUse a real example to explain the situation, what evidence changed your view, the action you took next and the result. This is suggested practice for discussing algorithm development and evaluation, not a claim about MediaTek’s interview process.

Sharpen your knowledge

Topics to prepare

Transformer and LLM inference
The posting explicitly requires understanding model graph structure, operators and inference behaviour. Trace a small transformer’s computation graph, identify its main operators, and explain how it produces successive tokens.
Alternative sequence architectures
The role names RNNs, Mamba and Linear Attention alongside transformers. Compare how each handles sequence history and discuss the implications for inference computation and memory.
Neural network optimization
Quantization, pruning and low-bit techniques are listed responsibilities. For a small model, explain what each changes and how you would check the effect on accuracy, model size and inference speed.
Edge AI and operator compatibility
The role includes APU optimization and requires familiarity with mobile or embedded frameworks. Review a model graph and identify which operators and data movements you would investigate before attempting on-device inference.
Multimodal and domain-specific algorithms
The posting covers language and multimodal models plus vision, voice and NLP. As preparation, sketch one small multimodal pipeline and explain its inputs, model components and outputs; this is a revision exercise, not a stated project requirement.

Skills to revise

Python and PyTorch model work
Both are listed programming requirements. Implement a small attention model in PyTorch, run an inference pass, and inspect tensor shapes and operators at each stage.
C/C++ implementation
C/C++ is explicitly required. Implement a simple neural-network operator, test its output against a Python reference, and explain any correctness or performance trade-offs.
TFLite and ExecuTorch inspection
The posting requires familiarity with both on-device frameworks. Inspect a small model in each, run inference where feasible, and note its inputs, outputs and operator support.
Optimization experiments
The role includes quantization, pruning and low-bit optimization. Apply one of these to a small model and record the change in model size, output quality and inference time, stating the measurement conditions.

Your preparation checklist

Map the stated requirements to evidence
Make a one-page table covering transformer and LLM inference, TFLite or ExecuTorch, and Python, PyTorch and C/C++. Note a project, exercise or concept you can discuss for each; leave gaps explicit rather than implying experience.
Prepare a model project walkthrough
Choose one relevant project and explain its architecture, training approach, results and your contribution. Trace an input through key graph operators to the output. If you lack a project, prepare a clearly labeled conceptual walkthrough instead.
Practice an architecture comparison
Sketch how a transformer such as LLaMa differs from an RNN or state-space approach such as Mamba, including likely inference and memory tradeoffs. Treat these as discussion points, not claims about MediaTek's implementation.
Work through an edge-optimization exercise
Draft a hypothetical mobile or embedded model plan using quantization, pruning or low-bit techniques. Specify what you would measure—such as accuracy, latency and memory—and how you would check operator support before considering APU optimization.
Prepare an on-device framework example
Outline or demonstrate a small PyTorch model's path to TFLite or ExecuTorch inference. Record any conversion or operator issues you actually encounter; do not claim device deployment unless you have tested it.
Rehearse a focused coding explanation
Prepare a short Python/PyTorch example and a C or C++ implementation sketch for a relevant tensor operation. Be ready to explain correctness checks and performance choices.
Write questions for the interviewers
Ask which model areas are priorities, what on-device constraints guide APU algorithm work, how optimization is evaluated, and where design automation or on-device learning fits the role.

Availability last checked: 4 Nov 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…