Most portfolios describe the work. I wanted mine to run it.

Every API route on this site — the chatbot, the JD-fit checker, the GitHub and LeetCode integrations — is instrumented. Each request records its latency into a capped list in Redis, and a /api/metrics endpoint computes p50/p99/avg on read. There's a /api/status route that does live dependency health checks and surfaces the deploy SHA and region.

Why bother

When a hiring engineer asks "can you build backends?", a paragraph is weak evidence. A live dashboard showing real latency percentiles of the very page they're looking at is strong evidence. The medium is the message.

The one rule: never fake the numbers

If Redis isn't configured, the dashboard shows a clearly-labelled sample state instead of inventing traffic. Streaming endpoints label their latency as time-to-first-byte, not full-stream. Cache metrics are framed as "origin executions", because the app layer genuinely can't see CDN hits.

Honest instrumentation beats impressive instrumentation. The moment a reviewer catches one fabricated metric, they distrust all of them.

That constraint — measure only what you can actually observe — is the same discipline that keeps production dashboards trustworthy.