50+services moved to ARM64
Zero downtime, about 25-30% lower compute cost.
Zero downtime, about 25-30% lower compute cost.
A custom OpenTelemetry Java agent sampler. The trace flood was about 60 GB every 6 hours, and the backend stopped filling up.
2 bare-metal (provider-managed control plane), 5 EKS with Karpenter. About 100 nodes and 160 services.
| NAME | STATUS | DESCRIPTION | LINK |
|---|---|---|---|
| cyclops | Running | Kubernetes controller (Go, kubebuilder) that emails or posts to Slack when a cert-manager Certificate is expired, never issued or overdue for renewal. Beta, v0.1.0-beta.8. | page · repo |
| otel-pattern-sampler | Running | Custom OpenTelemetry Java agent sampler. About 95% less trace volume, ended a 60 GB-every-6-hours trace flood. | repo |
| opentelemetry-helm-charts | Merged | opentelemetry-kube-stack chart: allow instrumentation.java.extensions in values.schema.json, so Java agent extensions pass Helm schema validation. Merged upstream (#2377). | PR #2377 |
| fluent-bit | Merged | Install script fix, merged upstream into fluent-bit. | commit |
A bash deploy script stuck in an infinite loop, fixed without a restart using gdb and setenv on the live process.
How a log travels from the app that writes it, through an agent like Fluent Bit, to the log store.
What observability is, what its three signals (logs, metrics and traces) are, and why we need it.
# [at] and [dot] on purpose: old-school spam-proofing, in homage to the ones who came before us
try: help · neofetch · man saksh · sudo
-- below: the longer version --
tl;dr? I love computers.
Long Answer: Managing systems at scale is a puzzle. What I like most is the hunt for the root cause of a problem, and the satisfaction of fixing it. I enjoy building systems that are reliable, scalable, and maintainable. I also just really like talking tech, hmu if you want to chat about anything from the kernel to the cloud.
replay of this page's own connection. timings: your browser. edge info: cloudflare.
contained a cloud credential compromise as first responder, then led a company-wide ssh key rotation and a move to keyless access. zero production downtime.
built solo a dual-arch jenkins pipeline: amd64 and arm64 docker builds run in parallel on native machines, then buildx merges them into one multi-arch tag. rolled 50+ services over with zero downtime, ~25-30% cheaper at the same scale and workload.
a service-critical 100 ms outbox check kept spamming traces and could not be turned off. about 60 GB every 6 hours, and elasticsearch came close to hitting its disk watermarks. i wrote an opentelemetry agent extension that drops just that pattern and leaves every other trace alone, and the same 60 GB now takes 4-5 days to arrive (otel-pattern-sampler).
7 kubernetes clusters (2 bare-metal, 5 eks), ~100 nodes, ~160 services.
fixed broken hairpin tcp on bare metal (the provider lb sent no proxy protocol header) with coredns rewrites.
workers catch sigterm and shut down gracefully, so long async tasks survive deploys and scale-downs.
600-vu load test found saturation at 26% of target load and 32x span amplification.
traced an elasticsearch translog runaway (24 GB translog vs 7.4 GB segments) across 49 shards.
merged upstream fixes in opentelemetry-helm-charts and fluent-bit.
wrote otel-pattern-sampler, an opentelemetry java agent extension that drops one noisy trace pattern and leaves every other trace alone.
built cyclops in go + kubebuilder: a controller that emails or posts to Slack a report of cert-manager certificates that are expired, never issued or overdue for renewal (beta).
first-author paper on cultural bias in LLMs, springer LNCS, PReMI 2025.
co-author: benchmarking llm powered chatbots, methods and metrics, arxiv 2023.
cyclops (beta, released): cert-manager renews your tls certificates automatically, until it silently doesn't. cyclops is the backstop: a controller that sends a report of certificates that are expired, never issued or overdue for renewal, and stays quiet when everything is healthy.
email via ses or smtp, working end to end on a local kind cluster. webhook, welcome message and helm chart are next.
go + kubebuilder + controller-runtime.
enveu: aug 2025 – present
careers360: dec 2024 – aug 2025
first-author paper on cultural bias in LLMs, springer LNCS, PReMI 2025
co-author: benchmarking llm powered chatbots, methods and metrics, arxiv 2023