We’re looking for a Machine Learning Engineer to own and push the limits of model inference performance at scale You’ll work at the intersection of research and production—turning cutting-edge models into fast, reliable, and cost-efficient systems that serve real users.
This role is ideal for someone who enjoys deep technical work, profiling systems down to the kernel/GPU level, and translating research ideas into production-grade performance gains.
Optimize inference latency, throughput, and cost for large-scale ML models in production
Profile and bottleneck GPU/CPU inference pipelines (memory, kernels, batching, IO)
Implement and tune techniques such as:
Quantization (fp16, bf16, int8, fp8)
KV-cache optimization & reuse
Speculative decoding, batching, and streaming
Model pruning or architectural simplifications for inference
Collaborate with research engineers to productionize new model architectures
Build and maintain inference-serving systems (e.g. Triton, custom runtimes, or bespoke stacks)
Benchmark performance across hardware (NVIDIA / AMD GPUs, CPUs) and cloud setups
Improve system reliability, observability, and cost efficiency under real workloads
Strong experience in ML inference optimization or high-performance ML systems
Solid understanding of deep learning internals (attention, memory layout, compute graphs)
Hands-on experience with PyTorch (or similar) and model deployment
Familiarity with GPU performance tuning (CUDA, ROCm, Triton, or kernel-level optimizations)
Experience scaling inference for real users (not just research benchmarks)
Comfortable working in fast-moving startup environments with ownership and ambiguity
Experience with LLM or long-context model inference
Knowledge of inference frameworks (TensorRT, ONNX Runtime, vLLM, Triton)
Experience optimizing across different hardware vendors
Open-source contributions in ML systems or inference tooling
Background in distributed systems or low-latency services
Real ownership over performance-critical systems
Direct impact on product reliability and unit economics
Close collaboration with research, infra, and product
Competitive compensation + meaningful equity at Series A
A team that cares about engineering quality, not hype
We enable serverless inference via our GPU orchestration and model load-balancing system. We unlock fine-tuning by enabling organizations to size their server fleet to throughput needs, not number of models in the catalogue.
See it in action on our public cloud, which offers inference for 10k+ open weight models.
