Member of Technical Staff - Inference & Hardware Optimization

Albs is an AI research lab building real-time multimodal intelligence for machines, enabling them to see, hear, reason, and interact. We treat model architecture, inference, and runtime as one system. Our purpose-built models and optimized runtimes unlock the full potential of each device within defined limits for hardware cost, power consumption, and response time. Companies can adapt our technology to their own machines without building the underlying AI from scratch.

We founded Albs at the intersection of LLM architecture research and on-device engineering, and we are currently in stealth, but well funded, with dedicated compute for large-scale training and experimentation. Publishing is a core part of our research culture, and we contribute our work to top venues. We share more details about the company, the team and our backing in the first conversation.

This is a Staff / Senior IC role. We are looking for experienced engineer, typically with several years of industry experience or an equivalent track record. The exact scope of each role depends on your background: some people go deep on one part of the stack, others shape the technical direction of a whole area. We agree on scope together with you during the interview process.

We welcome applications from all qualified candidates, regardless of gender, age, ethnic origin, religion, disability or sexual orientation.

• Work across our inference stack end to end: execution engine, memory planning, scheduling, and the low-precision kernels our models actually run on.

• Work directly with the research team on architecture and quantization decisions, and bring latency, memory and power requirements into the model design.

• Make real-time multimodal inference fast: streaming, KV cache management, time-to-first-token, and running speech, vision and language components side by side.

• Squeeze the last few percent out of every target: operator fusion specific to our architecture, cache-aware data layouts and threading. Build on existing runtimes where they serve us, and drop down to SIMD, GPU or NPU kernels when profiling shows that's where the remaining performance is.

• Bring up new hardware targets, from ARM CPUs and mobile GPUs to vendor NPUs, as we and our customers expand.

• Build and extend our benchmarking harness: per-layer latency profiling, memory traces, power measurement and accuracy checks across all the hardware we support

• You have several years of hands-on experience in performance engineering for numerical or ML workloads, in industry or research.

• You write strong systems code in C++, C or Rust.

• You have optimized ML workloads on constrained hardware at the kernel or runtime level, with hands-on experience in SIMD intrinsics (NEON, AVX), CUDA, or vendor NPU SDKs.

• You understand low-precision inference and quantization, and how they trade off speed against model accuracy.

• You are comfortable reading profiling traces and vendor documentation, and working out where the cycles went.

• Bonus: experience with ML compilers, streaming inference, or speech and vision models on device.

We are a small, focused team of experts. You would join early, work directly with the founders, and help shape how we build. Fast iteration, short lines of communication, and in-person discussion matter a lot to us. Our culture is built around the office in Freiburg, Germany, with a default of three days a week on site. Alternatively, you can work remotely and join us on a regular cadence.

We will discuss what works best for you during the interview process.

We hold ourselves to a high standard: the research must be rigorous, the understanding deep, and the product well crafted. Ideas…