Selected work

Whatwashired,whatchanged.

Six engagements with measured outcomes. Each one is a slice of the same operating philosophy: deliberate boundaries, tested failovers, observability tied to user journeys.

01·Fintech·14 weeks

Rebuilding the platform under a payments company without slowing the roadmap

Cut deploy time from 38 minutes to under 9, reduced cluster spend by 31%, and got the team out of a quarterly upgrade panic.

AWSEKSKarpenterArgoCDTerraformGitHub Actions

Outcome

Mainline build time: 38m → 9m
Monthly cluster spend: -31%
p95 checkout latency: -22%
On-call pages / week: 47 → 6

02·B2B SaaS·10 weeks

Standing up a platform team where there wasn't one

Delivered a working internal developer platform, paved-path service template, and hired the two engineers who own it now.

GCPGKEBackstageCrossplaneArgoCDGitHub Actions

Outcome

Time to provision a new service: 3 days → 22 minutes
DORA lead time for change: 4.1d → 9h
Engineer NPS on tooling: -12 → +41
Platform tickets per week: 60+ → 7

03·Healthtech·8 weeks

HIPAA-aligned cloud foundation for a clinical data startup

Cleared HIPAA technical safeguards review with their first enterprise customer's security team — on the first pass.

AWSEKSTerraformVaultCloudTrailGuardDuty

Outcome

Security review cycles: 1 (first pass)
Encrypted data stores: 100%
IAM roles with break-glass: All privileged paths
Audit log retention: 7 years, immutable

04·E-commerce·6 weeks (peak readiness) + ongoing retainer

Getting an e-commerce platform through Black Friday without a war room

Handled 7.2x the previous year's peak with a single sub-five-minute degradation, no all-hands incident, and a smaller bill than the prior year.

AWSEKSKarpenterRedisPostgreSQLDatadog

Outcome

Peak RPS handled: 7.2x YoY
Sev-1 incidents during peak: 0
Total peak-week cloud spend: -18% YoY
p99 cart latency at peak: <340ms

05·AI/ML Infrastructure·12 weeks

Bringing an AI infra bill back under control without slowing training

Cut monthly GPU spend by 44% while improving job throughput and shortening time-to-first-token on inference.

AWSEKSKarpenterKubeflowTritonPrometheus

Outcome

Monthly GPU spend: -44%
Training job throughput: +27%
Inference p50 TTFT: -31%
Spot interruption recovery: <90s checkpointed

06·Marketplaces·16 weeks

Real multi-region failover for a two-sided marketplace

Delivered a tested cross-region failover with documented RTO of 12 minutes and RPO under 30 seconds — and proved it in a live game day.

AWSEKSPostgreSQLpglogicalRoute 53Terraform

Outcome

Tested RTO: 12 minutes
Measured RPO: < 30 seconds
Failover game days: 4 (all successful)
Cross-region data sync lag p99: < 2.1s

Your project next

Ready to add a seventh row?

Start a conversation