About
I studied electrical engineering, so I think in signals: something goes in, passes through many stages, and comes out the other end. A web request works like that. So does a model being trained. My job is to keep every stage healthy, and when one breaks, to find the real cause instead of the loudest symptom.
I automate anything I have to do twice, and I write down what I learn so the next person doesn't have to learn it the hard way.
What I work on
From the hardware up to the model, one layer at a time. Each layer depends on the one below it, and most outages I have chased were really a problem one layer down.
-
roots
Machines and Linux
Where it starts: servers, disks, networks, and GPUs that need the right driver before they do any useful work.
- Linux
- NVIDIA GPUs
- networking
- shell
-
build
Infrastructure as code
Cloud resources written as code, reviewed like code, applied from a saved plan. Machine images built once and never patched by hand.
- Terraform
- AWS
- Packer
- Ansible
-
run
Scheduling, discovery and secrets
Deciding where each job runs, how services find each other, and how they get credentials without anyone pasting passwords. I also add features to open-source Terraform modules for this stack.
- Nomad
- Consul
- Vault
- Kubernetes
-
move
Data and messaging
Databases and streams, and the certificates that let them trust each other. Rotating those certificates without an outage is harder than it looks.
- PostgreSQL
- Kafka
- Redis
- mTLS / PKI
-
see
Observability
Metrics, logs and traces, so a problem shows up on a dashboard before it shows up in a complaint.
- Prometheus
- VictoriaMetrics
- Grafana
- OpenTelemetry
-
learn
Machine-learning pipelines
Training pipelines driven by config: one entry point, many models, each run tracked, scheduled on shared GPU machines like any other job.
- Python
- MLflow
- LightGBM
- GPU training
-
think
Language models
Running open models locally and building agents that route easy work to a local model and save the expensive model for the hard parts.
- MLX
- llama.cpp
- LoRA
- AI agents
How I work
- Root cause, not guessesIf the logs and the code are there, read them. "Probably" is not an answer.
- Simple beats cleverFewer moving parts means fewer 3 a.m. pages. A server I don't run can't go down.
- Plan, then applyEvery change is reviewed as a plan first. Nothing reaches production by surprise.
- Teach what I learnIf I can't explain it simply, I don't understand it yet.
Sites
How production systems really behave: databases, networks, the web, and the incidents that teach us.
The math of running language models on your own machine: memory, speed, and what actually fits.
Visual, sourced stories about people and ideas that shaped the world, starting with Gandhi.
The SRE and AI notes are private for now; your browser will ask for a login. Stories are open to everyone.