operational since 2008 · accepting workloads

AI-native platform engineering.
Deployed to your org.

I build and operate an internal AI agent platform on Kubernetes — agentic build triage, code search for agents, GitOps delivery — on top of 15+ years running production CI/CD at scale. This site is an instance of that work: the agent below is grounded in my real systems and cites its sources.

● agent@patelharsh.dev ~/experience
❯ how did the build-failure triage system work?
Harsh built a Jenkins Build Intelligence extension — 11 REST API tools plus an LLM skill encoding a failure taxonomy (compilation, test, dependency, infra, deployment, timeout, permission, quality gate). Agents resolve the build, gather log evidence, and emit a structured root-cause report. S1 · build-triage §3
→ run a Fit Report on your job description→ connect me to your Claude
$200K/yr
SaaS CI toolchain replaced in-house
10K+
pipeline runs / month operated
<1%
rollback rate at daily deploy cadence
120+ repos
migrated with <5 min DR RPO
18 yrs
in production, zero résumés harmed
what's in the box

Three capabilities, shipped and operated in production.

agent-platform · build-triage · code-search

Agentic Platform Engineering

Internal AI agent platform on EKS: LLM build-failure triage with a structured taxonomy, Zoekt code search wired into agent tooling, skills and extensions on the pi-coding-agent framework.

read the architecture →
jenkins · gh-actions · argo cd

Release & Reliability at Scale

Trunk-based workflows, GitOps delivery via Argo CD, canary/rollback strategy, and DORA dashboards leadership actually reads. Bi-weekly → daily releases with <1% change failure.

read the architecture →
eks · terraform · datadog

Cloud & IaC Foundations

Multi-cluster EKS operations, Terraform modules across prod/staging/dev, golden-image AMIs, cross-account secrets, fleet-scale patching and CVE remediation on Amazon Linux 2023.

read the architecture →
featured writing

Field notes from building a company brain.

Teaching an Agent to Triage: Automating First-Level Build Failure Analysis
Why "point an LLM at the logs" fails, the 11-tool Jenkins surface, and the taxonomy that turns red X's into ticket-ready root-cause reports.
2026 · 12 min
Code Search for Agents: Zoekt on EKS
Nightly indexing of 120+ repos on EFS, an arm64 webserver with an nginx sidecar, and why agents need codebase awareness before they need autonomy.
2026 · 10 min
I Productized Myself: The Architecture of This Site
RAG grounding, citation-mandatory prompting, an eval harness that blocks deploys, and a public MCP server. The portfolio as a working system.
2026 · 9 min
the api

Connect me to your Claude.

This site exposes a public, read-only MCP server. Add it to Claude and query my experience from inside your own workflow — search projects, pull the structured resume, or run a fit analysis against your open req.

MCP endpoint copied
# Claude → Settings → Connectors → Add custom connector
 
name: harsh-patel
url: https://patelharsh.dev/api/mcp
 
# then ask: "is Harsh a fit for this JD? <paste>"
fit report

Paste your job description. Get an evidence-linked analysis.

Requirement-by-requirement mapping against my actual shipped work — with citations, and an honest gaps section that tells you where the evidence is thin. The cover letter, inverted.

run a fit report →