Case studies
Four engagements, described from the work itself.
30TB electronic health records migration to Azure
Situation. A client's entire EHR platform had to move off on-premises
infrastructure into a new Azure region: 30TB of live healthcare data, where data loss was
not an option.
Approach. I migrated the estate with a mix of image-level, SQL, and
file-level backups, moving data over SFTP and image restores into the new region. Cutover
was phased, with validation gates between phases.
Result. All 30TB landed in Azure with zero data loss, verified by
SHA256 checksum comparison, phased cutover validation, and post-migration audits plus
smoke testing of system functionality.
Handover. The client team kept the full cutover and rollback
runbooks.
Local LLM serving: W4A16 AWQ on SGLang
Situation. Large-model inference on owned hardware where the data
stays in-house, at context lengths and concurrency consumer-grade setups cannot
reach.
Approach. I quantized the model to W4A16 AWQ and served it with
SGLang, running tensor parallelism across an NVLink-paired dual-GPU rig with weights and
KV state split across cards. MTP/NEXTN speculative decoding and context-pool sizing were
tuned under concurrent load. Benchmarks ran before and after every change, and
long-context configurations were probed past 90% of the target window before anything
shipped.
Result. +30% single-stream throughput from speculative decoding, with
96k-token parallel contexts coexisting with a realtime audio stack on the same
GPUs.
Handover. The benchmark-first methodology is documented and repeats
on every model-serving deployment I build.
75%+ Azure footprint reduction
Situation. An estate audit requested by the client's security
leadership surfaced unrestricted RDP reach into every VM, and a cloud bill that looked
wrong.
Findings. Not one resource carried a cost-saving reservation.
Numerous resources sat idle, and several were completely unneeded: compute, storage,
public IPs, VPN gateways.
Approach. RDP was locked down to the client's own addresses, then I
redesigned the environment and cut over to it completely, data and endpoints included.
Several months end to end.
Result. Azure footprint reduced 75%+ and monthly billing down by
thousands of dollars.
Handover. Cost and security reviews are now a standing part of how I
maintain client estates.
Scale-to-zero Azure Virtual Desktop
Situation. A financial-services client needed tightly controlled
access to internal records without paying for idle session hosts.
Approach. A three-host Azure Virtual Desktop pool: RBAC decides who
reaches what, FSLogix roams profiles on Azure Files, and host VMs scale to zero outside
working hours.
Result. Compute spend for the pool dropped by roughly $900 a month.
Users reported the environment performed far better than its predecessor: more
responsive, user friendly, and cost effective.
Handover. Built entirely from managed Azure services with the
configuration documented, so the client team operates it day to day.
