Help us build the fastest way to run intelligence.
We're building a high-performance inference stack from the kernel up, on our own GPU infrastructure.
Open roles
Real-time voice intelligence
Voice AI Engineer
Build and optimize real-time voice agents on top of Blaze's inference stack — the low-latency backbone for products that need to listen, think, and respond in milliseconds.
- Build and tune streaming pipelines for speech-to-text, text-to-speech, and full-duplex voice agents running on Blaze
- Drive down end-to-end latency across the voice stack, from audio capture to model response
- Work with the inference team to shape serving behavior for voice-specific needs like barge-in, interruption handling, and fast time-to-first-audio
- Evaluate and integrate new voice models and architectures as they emerge
High-throughput model serving
Inference Engineer
Own the serving layer that turns raw GPU throughput into fast, reliable tokens — the core of what makes Blaze, Blaze.
- Build and operate high-throughput, low-latency model serving on our own GPU infrastructure
- Tune batching, scheduling, and memory management for maximum tokens per second per GPU
- Design and maintain the routing layer that keeps every request on the pod with the most headroom and the warmest cache
- Diagnose and fix performance regressions across the stack, from the API down to the scheduler
GPU performance & custom kernels
LLM Kernels Engineer
Go deep on model architectures and GPU internals to squeeze more performance out of every chip we run.
- Write and optimize custom CUDA kernels for attention, matmul, and quantization paths used in production inference
- Profile kernels with tools like Nsight Compute and Nsight Systems, and turn findings into concrete speedups
- Understand model architectures deeply enough to know where a kernel change actually moves the needle
- Track new architectures — MoE, long-context, multimodal — and adapt kernels to support them efficiently
Kubernetes & LLM workloads at scale
Site Reliability Engineer
Keep a fleet of GPUs serving real-time inference traffic healthy, observable, and resilient at scale.
- Operate and scale Kubernetes clusters running latency-sensitive LLM workloads
- Build monitoring, alerting, and capacity-planning tooling tuned for GPU-bound inference — KV-cache usage, queue depth, throughput
- Design for failure: rolling upgrades, pod health checks, and graceful degradation under load
- Partner with the inference team to diagnose production incidents and land lasting fixes, not just patches
Don't see your role?
We're always open to talking to strong engineers. Send us a note with what you'd want to work on.
[email protected]