platform engineer · llm inference & mlops · melbourne
Ocha Cakramurti
I build and run inference infrastructure. As the founding infrastructure engineer at Heidi Health, I took an AI medical scribe from its first AWS account to five data-residency regions, and now run its self-hosted speech-to-text platform on GPUs.
production impact
- ~20%lower cost after replacing a managed speech-to-text vendor with self-hosted GPUs
- ~7xfewer errors than the vendor, with ~28% lower p95 latency
- 3.4xthroughput per GPU node after moving to MIG slicing
- 4 in 5 modata-residency regions launched: UK, Canada, US, EU
lab
all demos →GPU slicing calculator
How many model replicas fit on a GPU, and when MIG slices beat whole GPUs on cost per unit of throughput.
Inference fleet simulator
Autoscale a GPU inference fleet through a traffic day, then survive an on-demand capacity outage.
Gateway failover
Retries, failover and data residency across three regions behind an inference gateway.
experience
full cv →- 2025 - nowSenior Cloud Engineer, Engineering PlatformHeidi Health
- 2025Acting Engineering Manager, DevOpsHeidi Health
- 2022 - 2025Founding infrastructure engineerHeidi Health
- 2015 - 2021Risk, audit and dataEY, PwC, Cultural Infusion
talks
- AWS APJ EKS RoundtableTokyo · Jul 2026 · invite-only
ECS to EKS migration and multi-region GPU inference under data localisation
- AWS re:Invent 2025Las Vegas · Dec 2025 · lightning talk
How Heidi Health is leveraging GenAI to transform the global healthcare industry watch →
- AWS Startup Dev DayMelbourne · Aug 2025 · with Kerry Zhang
MongoDB customer session: the DocumentDB to Atlas migration