Inference Optimization — Member of Technical Staff

Construct Labs; Berlin; Vor Ort; EUR 120.000 - 180.000; Vollzeit

Zusammenfassung

Construct Labs in Berlin is seeking an Inference Optimization — Member of Technical Staff to push GPU efficiency for deployed models. You will identify novel hardware-informed quantization methods, write and tune Triton and CUDA kernels, and locate bottlenecks in compute, memory, and communication.

Great candidates may come from ML systems, compilers, mathematics, physics, competitive programming, security, or HPC, and share a habit of deep reasoning and a passion for optimizing end-to-end

Qualifikationen

• Experience with model quantization techniques and hardware-aware optimization.
• Ability to profile and optimize GPU workloads across compute, memory, and communication.
• Strong foundation in machine learning systems and performance engineering.

Aufgaben

• Find novel hardware-informed quantization methods across model architectures.
• Write and tune Triton and CUDA kernels.
• Identify bottlenecks across compute, memory, and communication.
• Efficiently serving LoRAs and specialized weights for model adaptation.

Kenntnisse

Quantization methods

GPU profiling

Performance optimization

ML systems

Tools

Triton

CUDA

Profiling tools

Jobbeschreibung

Back to careers

Inference Optimization — Member of Technical Staff

Berlin, Full-time, In person

Who Are We

Models should learn from what happens after deployment. We're building the loops that make this possible.

We will be the default inference provider for specialized tokens, running models that continuously adapt to each customer and workload. That means serving many continuously adapting models with frontier-level performance and attractive economics.

The Work

• Finding novel, hardware-informed quantization methods across model architectures


• Writing and tuning Triton and CUDA kernels


• Finding bottlenecks across compute, memory, and communication


• Efficiently serving LoRAs, learned memory, specialized weights, and other forms of model adaptation

This is capture-the-flag for GPU efficiency: profile the system, find something everyone else missed, and prove the gain on real workloads.

Great candidates might come from ML systems, compilers, mathematics, physics, competitive programming, security, or HPC. The common thread is strong first-principles reasoning, an obsession with efficiency, and a habit of going deep on hard problems for the fun of it.

We work together in person from our office in Berlin.

Interview process

• Two technical interviews


• An onsite interview at our Berlin office