Maksim Miller

I'm a senior engineer working on LLM inference infrastructure: making large models serve fast, reliably and cheaply on GPUs.

At NVIDIA I own a serving tier for hosted NIM: 40+ models for 900+ customer teams, about 4.2B tokens a day on ~180 GPUs. The hard part is economics rather than raw throughput. Reworking model placement with scale-to-zero, NVMe weight streaming and multi-LoRA cut GPU-hours per million tokens by 46%. On the latency side, FP8, prefix caching, cache-aware routing and chunked prefill brought p95 time-to-first-token from 920 ms to 410 ms.

I also built the release gate every model, quantization and engine upgrade passes before reaching customers: 1,400 cases scored by an LLM judge calibrated against human labels. I own embedding and reranker serving too, where I cut unsupported RAG claims from 11.4% to 3.9%. Alongside the engineering I plan capacity for the tier, mentor engineers and take on-call.

Before NVIDIA I spent two and a half years on search and ranking at Yandex Market: neural ranking and semantic retrieval, and p99 latency from 210 to 95 ms at 6.8K RPS.

I'm interested in inference optimization, serving economics and evaluation. I'm looking for remote roles on teams where LLM serving is the core product.