AI Workloads Engineer

AI Workloads Engineer

Belgrade (Serbia)

Hybrid

Job Openning

Copy link

Belgrade (Serbia)

Hybrid

• Location: Beograd, Serbia
• Work mode: Hybrid, 2 days/week in the office

Adaptiq, on behalf of NextSilicon, is looking for a AI Workloads Engineer.

About NextSilicon:

NextSilicon is reimagining high-performance computing (HPC) and AI through accelerated compute solutions powered by intelligent adaptive algorithms. Its novel software-defined hardware architecture is designed to significantly accelerate supercomputers and advance performance across HPC and AI workloads.

The company is guided by three core values: professionalism, unity, and impact, bringing together engineers who care about quality, collaboration, and meaningful technological progress.

About the Role

This is an experienced, hands-on software engineering position in a small AI workloads team. You will build Python-based inference and performance-modeling infrastructure, run and adapt real workloads, and turn measurements into findings that hardware, software and business stakeholders can use.

The scope includes both existing and next-generation hardware. You will have room to take work through implementation, profiling and reporting rather than handling only narrowly defined tasks. Prior chip expertise is not required; the team provides hardware onboarding. The role is based in Belgrade, Serbia, with hybrid working: normally three days in the office and two at home. A two-day office arrangement may be possible for someone living farther away. Regular office attendance is required; the role is not fully remote.

Key Responsibilities

• Build and maintain features for inference workflows and performance models, contributing code through pull requests.
• Model end-to-end inference behavior across simulation and execution on hardware.
• Run and adapt open-source AI models to investigate how they perform on current or planned chips.
• Instrument and profile workloads; analyze latency, throughput, batching and request-level statistics.
• Investigate workflows involving assisted decoding, dynamic batching and key-value cache use.
• Use SDK and framework-integration tools to identify full-stack bottlenecks.
• Connect tools that do not already work together, and prototype runtime components or kernels when a workflow requires them.
• Work with compiler, runtime and hardware design teams on potential improvements.
• Present performance estimates and findings clearly to technical and business stakeholders.

Required Competence and Skills:

• At least 4 years of software engineering experience, with substantial hands-on development.
• Strong, current Python skills and the ability to understand and review the code you work with.
• Strong PyTorch development experience.
• Practical experience with LLMs or relevant AI systems beyond calling a hosted API, such as examining or adapting open-source models.
• Understanding of transformer-based models and attention, with familiarity with inference concepts such as KV caches, batching or assisted decoding.
• Sound problem-solving skills and the ability to explain your own work, decisions and results to different audiences.

Nice to Have:

• Experience with distributed inference, model parallelism, NCCL collectives, dynamic batching systems or MLPerf Inference.

• Custom Kernel development with CUDA, Triton or similar tools.

• Hardware-aware optimization, simulation, workload modeling, or compiler and runtime profiling.

• C++ exposure, particularly where software behavior is tied to latency, dataflow or memory access.

• Thank you for your

• interest in us.

• We will reach out soon!

• Thank you for your

• interest in us.

• We will reach out soon!