Principal Machine Learning Engineer

Unity Technologies SF•Mountain View, CA

About The Position

We are building the next generation of AI-driven game experiences, running generative models on-device, right where the players are — on phones, tablets, laptops, and desktops. Our games run inside a modern, browser-native runtime (built on technologies such as WebGPU and WebNN), so the models that power these experiences must be deployed and accelerated entirely within that runtime. As our Principal Engineer for On-Device AI Inference & Systems, you will be the foremost engineering authority on taking state-of-the-art multi-modal models (transformers and diffusion networks) and making them run fast, small, and reliably within that runtime, fully integrated into a production game engine. This is a deeply hands-on, high-impact engineering role. You will own the inference and integration stack end-to-end — from the moment a trained checkpoint leaves research, through export, optimization, and kernel-level tuning, to a shipped feature running inside the engine at interactive frame rates within a fixed memory and power budget. You will set the engineering standards, drive the architecture of the runtime and integration layers, and mentor a team of senior and mid-level engineers. Your work directly determines the latency, quality, memory footprint, and battery profile of AI features experienced by players worldwide. This role is for an engineer who is energized by the gap between a research model and a shipping, AI-based product. If you love profilers, frame captures, op-fusion, and shaving milliseconds and megabytes, this is your role.

Requirements

8+ years in software/ML engineering, with at least 4 years focused on on-device / edge inference or real-time, performance-critical systems.
Proven production deployment of transformer- and/or diffusion-based models (e.g., ViT, Stable Diffusion) on mobile, desktop, or embedded hardware — shipped, not just prototyped.
Hands-on experience deploying models through WebGPU — e.g., ONNX Runtime Web (WebGPU EP), Transformers.js, WebLLM, or TensorFlow.js — including writing/tuning WGSL compute shaders and working within WebGPU's adapter, device-limits, and binding model. Equivalent deep experience with a native GPU/compute API plus a clear path to WebGPU will also be considered.
Hands-on expertise with at least one major inference runtime (ONNX Runtime / ORT Web, CoreML, TFLite, ExecuTorch) and deep understanding of operator fusion, memory layout, and runtime scheduling.
Low-level performance engineering: strong command of at least one GPU/compute API — WebGPU/WGSL, Metal, Vulkan, D3D12, or CUDA — and the profiling tools to go with it. You can read a frame capture and a kernel trace and know where the time and memory go.
Working knowledge of model-optimization techniques — quantization (INT4/INT8/FP16), weight sharing, pruning, and distillation — and the practical judgment to apply them to hit latency and memory budgets. You don't need to be a research expert in these methods; you need to use them effectively as engineering tools.
Strong understanding of target hardware: mobile SoCs (Apple Neural Engine, Qualcomm Hexagon/Adreno, ARM Mali) and desktop/laptop GPUs (Apple Silicon, NVIDIA, AMD, Intel) — and how to target each for peak throughput.
Proficiency in the core languages of a browser-native runtime — TypeScript/JavaScript and WGSL — plus solid Python for export pipelines and training-side tooling.
Working fluency with the models you deploy — enough to read an architecture, modify it for deployment, and reason about accuracy trade-offs.
Track record of technical leadership: setting engineering direction, influencing cross-functional partners, and growing engineers.

Nice To Haves

Experience shipping world-model, neural-rendering, or real-time generative pipelines (NeRF, 3DGS, real-time diffusion, or similar) on device.
Deep game-engine or real-time-graphics background (Unity, Unreal, or a custom engine; Metal/Vulkan/D3D/OpenGL ES render pipelines) — especially integrating compute workloads alongside a renderer.
Contributions to open-source ML inference frameworks, runtimes, or GPU/compute libraries — especially in the WebGPU ecosystem (Dawn, wgpu, ORT Web, Transformers.js, WebLLM).
Familiarity with the WebGPU specification and its evolving compute features (subgroups, FP16/shader-f16, timestamp queries) and the trade-offs of running heavy diffusion workloads in the browser/web runtime.
Familiarity with compiler stacks (MLIR, TVM, IREE, XLA) for custom kernel generation and graph optimization.
Experience with on-device benchmarking infrastructure, performance-regression CI, and large device-farm matrices.

Responsibilities

Own the end-to-end optimization pipeline: model export, graph transformation, operator fusion, memory-layout planning, and hardware-specific kernel tuning across NPU, mobile GPU, and desktop/laptop GPU.
Make authoritative decisions on quantization (INT4/INT8/FP16), weight sharing, structured/unstructured pruning, and knowledge distillation to hit hard latency, memory, and power budgets — and validate them against quality bars.
Drive low-level performance work: write and tune WebGPU compute shaders (WGSL) and, where relevant, native kernels (Metal, Vulkan/SPIR-V compute, D3D12, CUDA); profile with browser and platform tools (Chrome/Dawn GPU traces, PIX, Instruments/Metal System Trace, Snapdragon Profiler, Nsight, RenderDoc), and eliminate bottlenecks at the op and memory-bandwidth level.
Apply efficiency techniques — dynamic resolution, token reduction, cross-frame caching/reuse, reduced-step diffusion samplers — as engineering levers to meet budgets on target SKUs.
Evaluate, select, and drive adoption of WebGPU-targeted inference runtimes (ONNX Runtime Web, Transformers.js, WebLLM, TensorFlow.js) alongside native options (CoreML, ONNX Runtime, TFLite, ExecuTorch) — and extend or build runtime/glue code where off-the-shelf options fall short of our diffusion workloads.
Design and own the integration between the ML runtime and the game engine: real-time scheduling, threading, memory pooling, zero-copy buffer sharing between the inference path and the render path, and frame-budget management alongside the renderer.
Architect inference systems that handle diverse inputs — images, text, primitives, metadata — and produce pixel-level outputs with real-time performance, robust to the messy realities of production (cold starts, thermal throttling, device fragmentation, backgrounding).
Build the supporting engineering: model packaging and asset pipelines, on-device fallbacks and SKU-aware capability tiers, crash/quality telemetry, and automated on-device benchmarking in CI.
Partner closely with research scientists to turn novel architectures into implementations that are deployable, debuggable, and fast on device.
Provide the feedback loop back into research: surface hardware constraints, op-support gaps, and cost models early so model design and deployment converge.
Track breakthroughs in efficient inference (efficient attention, distillation, reduced-step diffusion) and assess them pragmatically: what actually moves latency/memory/power on our target devices, and what is worth the engineering cost.
Lead and mentor a team of engineers; set engineering best practices, code-review standards, performance-regression gates, and on-device benchmarking methodology.
Champion a culture of measurement: define and enforce KPIs for latency, quality, memory, and power, and ensure they are tracked rigorously across the device matrix.
Partner with platform engineers, product managers, and runtime teams to align ML capabilities with device-SKU constraints and product roadmaps.