Member of Technical Staff (AI Inference Engineer)
At a glance
Senior ML Platform & Ops role at Perplexity. San Francisco · full-time · $220,000–$485,000 base.
aimlroles summary, based on the employer's posting.
What you'll do
- Add transformer retrieval, text-generation and multimodal models to inference systems, handling weights, scheduling and KV-cache management
- Convert internal CUDA kernels to CuTe DSL for GB200 systems and Vera Rubin racks
- Develop the Rust inference server and tune bottlenecks across network ingress, batching and GPU kernels
- Set up dashboards, alerts and automated recovery, then investigate production incidents
What you bring
- At least three years of software engineering, including substantial ML inference or high-performance systems work
- Deep GPU programming and performance skills with CUDA, Triton, CUTLASS or a comparable tool
- Experience with at least one deep learning framework, such as PyTorch, JAX or TensorFlow
- Knowledge of GPU architecture, LLM designs and inference techniques such as quantization or speculative decoding
Who this fits
This senior individual engineering role suits you if you can move between Rust, Python and GPU kernels, and carry work from research into production. You should be comfortable operating distributed systems under real load and finding a path forward independently.
From the employer
We build and run the inference engine behind every Perplexity query and deploy dozens of model architectures at scale with tight latency and cost budgets. Our stack is Rust, Python, CUDA, and CuTe DSL - and we need another engineer to join us.
What you will work on
Examples of real work the team does:
New models support. Support transformer-based retrieval, text-generation, and multimodal models in our inference infrastructure, from weight loading, request scheduling and KV-cache management to support in API Gateway.
GPU kernels migration to CuTe DSL. Port our in-house CUDA kernels to NVIDIA's CuTe DSL so they run on GB200 today and are portable to Vera Rubin racks tomorrow.
Rust-native serving runtime. Develop our internal Rust-based inference server to solve all Python pains and keep up with rapidly growing traffic.
Performance optimisation. Profile and fix bottlenecks from network ingress through continuous batching and GPU kernel interleaving.
Reliability and observability. Build dashboards, alerts, and automated remediation so we catch regressions before users do. Respond to and learn from production incidents.
Who we're looking for
Deep experience with GPU programming and performance work (CUDA, Triton, CUTLASS, or similar). Any other deep systems programming experience is a plus.
You understand modern LLM architectures and are able to bring them up reliably in a production environment.
You've built and operated production distributed systems under real load - ideally performance-critical ones.
Comfortable working across languages and layers: Rust for the serving runtime, Python for model code, CUDA/CuteDSL for kernels.
You own problems end-to-end. You can read a research paper on Monday, write a kernel on Wednesday, and debug a production incident on Friday.
Self-directed. You do well in fast-moving environments where the path forward isn't laid out for you.
Good if you touched any of
ML compilers and framework internals: PyTorch internals, torch.compile, custom operators.
Distributed GPU communication: NCCL, NVLink, InfiniBand, RDMA libraries, model/tensor parallelism.
Low-precision inference: INT8/FP8/FP4 quantization, mixed-precision serving.
Profiling and debugging tools: Nsight Compute/Systems, CUDA-GDB, PTX/SASS analysis.
Container orchestration: Kubernetes, GPU scheduling, autoscaling inference workloads.
Qualifications
3+ years of professional software engineering experience with meaningful work on ML inference or high-performance systems.
Familiarity with at least one deep learning framework (PyTorch, JAX, TensorFlow).
Understanding of GPU architectures (memory hierarchy, warp scheduling, tensor cores).
Understanding of common LLM architectures and inference optimization techniques (e.g. quantization, speculative decoding, prefill-decode disaggregation).