Cloud & distributed systems: AWS consulting
A bill that grows faster than your traffic, APIs that slow down at peak, failures that spread from one service to the next: most infrastructure problems have a known shape, and the hardest have been studied for decades. I design from those principles on AWS, GCP, Azure or Kubernetes, weigh every choice for cost, performance, security, complexity and developer friction, and pressure-test each design with frontier models before anything is built.
What I build for
Graceful overload
Demand will outrun capacity sometimes. I give every part of the system a limit and decide in advance what happens when it's reached: queue the work, slow it down or turn some of it away.Failures that stay small
One service going down shouldn't take the others with it. I bound every call with a timeout, cap retries, and stop calling a service that's down until it recovers, so a brief failure stays brief.Room to grow
Growth should mean adding machines. I split the work and the data so load spreads evenly as you grow, and so no single customer can slow down everyone else.Fast at peak
A system has to be fast in its busiest hour and for its biggest customers. I load-test at the traffic you expect next and fix whatever slows the heaviest requests first.Infrastructure your team owns
I write every environment down as code, review every change before it ships and keep test environments that match production, so your team can change the system without me.From client work
All case studies →Questions
Our app slows to a crawl every time traffic peaks. Do we have to rebuild it?
Usually not from scratch. One IoT platform in my case studies averaged 12-second API responses and slowed under peak demand. I led a three-person team that stayed on Lambda and SQL Server and rebuilt the foundations underneath: environments to test in, infrastructure in code, pooled database connections, and queues that absorb traffic spikes rather than passing them straight to the database. Average response time fell to 150 ms, and load tests showed room for twice the devices on the same infrastructure.
Our cloud bill is growing faster than our revenue. Where is the money going?
Usually into a few common choices that cost more as you grow. Provisioned concurrency, the usual fix for slow Lambda starts, bills around the clock: 100 copies of a function kept warm at 1 GB cost about $1,080 a month before they serve a request. API Gateway’s REST API charges $3.50 per million requests where its HTTP API charges $1.00, and for steady traffic, paying per request stops saving money at all. Sometimes it’s the database: moving one IoT platform’s analytics from SQL Server to ClickHouse Cloud cost $2,200 a month less. I measure what each part really uses before changing any of it.
We’re scared to change anything in production. How do we make changes safe?
Give every change somewhere to be tried first, and a way back. We wrote the existing production setup down as code for an IoT platform, rather than building it again from scratch, and used that code to create development and staging environments to test in. From then on, every change shipped with zero downtime: new features behind flags, each deploy and its rollback tested outside production first, and rollouts in stages, watched at each step.
Only one person really understands our infrastructure. What happens if they leave?
Not much, once it’s written down. I write every environment down as code, so anyone on the team can read it, review a change to it and build a copy of it. When we finished on the IoT platform, its own team had documentation for every system, recorded walkthroughs of key operations and troubleshooting, hands-on training, and monitoring dashboards with alert thresholds and runbooks, so it could run the system after we left.
We get errors at peak, but our database isn’t even busy. What’s going on?
Usually it’s the connections to the database that run out. On Lambda, every copy of a function running at once opens its own connection, so 1,000 concurrent invocations can mean 1,000 connections, more than most databases can handle. RDS Proxy pools them, but it often has to hold a connection for one caller for the whole session, especially with SQL Server; AWS calls this pinning. Capping Lambda’s concurrency at what the pool can serve gives steadier performance and fewer timeouts. On the IoT platform, we pooled its connections with RDS Proxy, managed concurrency and changed transaction handling to stop needless locks.
Tell me what you're building
A free 30-minute call with me, not a sales team. Bring the diagram or the prototype. My engagements run long and personal, so I'd rather know we fit first. Either way, you'll get a Slack Connect invite for later questions.
Book a call