GPU Inference Autoscaler on Kubernetes
A latency-aware GPU autoscaler for LLM and vision inference with batching, queueing, and spot fallback.
KubernetesPythonAWSCI/CD

Highlights
- Continuous batching and priority queueing tuned per model.
- SLO-driven autoscaling with spot + on-demand blending.
- Warm-pool preloading to hide cold-start on large models.
Outcomes
- Cut inference cost per 1k tokens by 47% at matched p99 latency.
- Sustained 99.95% availability across bursty production traffic.
Stack
- Kubernetes
- Python
- AWS
- Docker
- CI/CD