Engineering write-ups on AI/ML, LLM inference, and building production software
How we built an autonomous clinical coding pipeline with Google MedGemma 27B on MIMIC-IV: doubling primary diagnosis accuracy to 40%, cutting note latency by 63%, and solving the 97,441 ICD-10-CM code parametric bottleneck with two-stage canonical grounding.
19.3x higher throughput, 80ms vs 673ms P99 latency, and PagedAttention under the hood — here's the real engineering breakdown of why developer UX doesn't equal production scale.
From 54GB to 15GB, from 0 to 600+ tokens/second — here's the real engineering story behind serving Google's MedGemma 27B on a single GPU.