RanarTech Insights
Opinionated engineering notes from the RanarTech team. Production patterns, cost comparisons, and the boring details of shipping software that survives contact with reality.
Designing LLM applications that survive contact with production
Nine engineering patterns we use on every LLM project: prompt versioning, eval-first development, retrieval freshness, cost ceilings, and the ones that matter most.
Multi-region HA on AWS and GCP: the boring parts that matter
Most teams do not need multi-region. When you do, here is the order in which to build it: RPO/RTO math, multi-AZ, then multi-region, with real cost numbers.
Validator infrastructure that does not page you at 3 AM
Twelve years of running production validators distilled into the patterns that matter: HSM-backed keys, double-sign protection, sentry architecture, MEV-aware infra, and the runbooks you actually trust.
Zero-downtime deploys without Kubernetes
Most teams adopt Kubernetes for deploy orchestration they could get from nginx, a load balancer, and a shell script. Here is when k8s is right, when it is not, and what to use instead.
E2E test suites in CI that engineers actually run
The shape of a test suite that survives contact with a real engineering team: small enough to be fast, flake-budgeted, run on every PR, and trusted enough that on-call engineers follow its signal.
Architecture reviews that change a team's trajectory
What we look at first, how we interview, the antipatterns we find in week one, and how to write a review that people actually read and act on.
RAG versus fine-tuning: pick the boring one
After two dozen LLM projects, here is the actual decision tree for when to use retrieval, when to fine-tune, and when neither is the right answer.
How to size an LLM eval suite without losing a quarter to it
Concrete guidance on what to measure, how many cases you need, where golden sets end and synthetic data begins, and how to keep the suite trustworthy as it grows.
Multi-region HA: AWS versus GCP, in real numbers
A concrete cost and latency comparison for active-active multi-region deployments on AWS and GCP, with the hidden costs nobody quotes you in the sales deck.
The 23-item validator infrastructure checklist we use before we call it 'done'
Keys, monitoring, runbooks, upgrades, backups, security. The full checklist, with what we check, what we reject, and the order in which we build.
Zero-downtime deploys when you are not using Kubernetes
Practical patterns for production deploys without k8s: nginx with Consul, blue/green on plain VMs, traffic shifting with HAProxy, and the database migration rules that make any of them safe.