All work

Case study

Local LLM Deployment & Serving

Self-hosted LLMs on owned GPUs serving two production pipelines, cutting cost per document from ~$0.35 to ~$0.10.

Local LLMsQWENLlamaGPU InferenceKV CacheCUDADocker

Problem

Production pipelines depended on third-party LLM APIs: ongoing per-token costs, rate limits, and sensitive medical data leaving infrastructure.

Naive open-source serving wasted GPU memory and could not keep up with production load.

Solution

We deployed serving instances across RTX 4090s and an A100 80GB, wired into extraction and chatbot pipelines.

Serving was optimized with KV cache management, prefix caching, and request batching sized to GPU memory budgets.

Outcomes

Two self-hosted instances serve production traffic.

Inference cost per document dropped from ~$0.35 to ~$0.10; sensitive documents stayed in-house.

Have a similar challenge?

Start a project