Kumo

An achievement worth making legible.

Running real production workloads entirely on cloud free tiers is an infrastructure achievement. Kumo turns the raw usage reports behind it into a clear picture—and a portfolio piece, showcasing data storytelling, infrastructure optimization, and lean engineering in one small static site.

Five infrastructure panels. Plus the other line item.

The dashboard frames infrastructure around five questions, each answered by its own panel:

• Monthly Cost Actual spend leads—$0.00 for an all-free fleet—with the list rate beside it, so the value the free tier absorbs is visible, not hidden.
• Free-Tier Utilization A gauge per allowance: this month’s burn against the monthly budget.
• Cost Trend Per-provider daily spend—the chart that spots the anomaly first.
• Resource Usage A daily series is flat (always-on resources burn steadily), so this plots cumulative use since the 1st as a share of each allowance—the slope is burn rate. The axis zooms in early in the month; the 100% free-tier ceiling appears as usage approaches it.
• Infrastructure Yield What’s actually running for the cost: the fleet, its uptime, services, and backups.
• AI Usage The other line item: token volume, cache efficiency, and what the agents cost at published rates—on the AI usage view.

The other line item: AI.

Free infrastructure is half the cost story. The AI usage view shows the other half—token volume, cache efficiency, and what the agents cost at published per-token rates. Models without documented rates appear as unpriced volume, never as free: the same show-the-basis discipline as the $0.00 headline, applied to AI. That discipline extends to comparisons: each model's blended rate divides its estimated cost by its own tokens—workload mix and caching included—so it describes this window's usage, never a catalog price.

Outcomes: work, verified.

Cost says what the tools consume; the Outcomes panel says what got shipped. Every entry is a work unit verified against the sessions that produced it—deterministic checks plus an independent AI auditor, with ambiguous cases escalated to a human—covering things shipped, repaired, automated, and investigated. Only weekly rollups publish. Individual sessions, prompts, and titles never leave the machine, and weeks with fewer than three verified units stay private rather than hint at them.

What counts as an outcome? Six plain verbs, tested by effect rather than effort: something shipped (a new capability), repaired (broken made right), automated (recurring hands-on work became zero-touch), investigated (a question answered with evidence), documented (writing others can act on), or other.

Who checks these? A two-stage auditor: deterministic checks first—sessions attached, evidence present, dates sane—then an independent AI judge reads the evidence and approves or rejects with reasons. Anything ambiguous escalates to me for the final call, and my veto always wins.

Built with almost nothing.

The whole pipeline is shell and the sqlite3 CLI—zero runtime dependencies, no build step, running natively on the ARM hosts it watches. Usage lands in a single SQLite file committed to git, and that file is the backup: no database server, no separate backup infrastructure.

Charts are Chart.js, vendored into the repo and version-pinned—same-origin, no third-party CDN. The site is static, served by Cloudflare Pages. A daily timer runs ingest → generate → publish, and every push deploys. Adding a provider is additive, not a rewrite—each one a small module that normalizes into one schema.

Built for more than free.

Today’s fleet is all free, so the headline reads $0.00—but the pipeline already handles paid usage and overage. The moment real spend appears, the headline simply moves with it. More providers can drop in through the same model.

What it does, not how to reach it.

Kumo is an open portfolio dashboard. It shows what the infrastructure does, not how to reach it—no IP addresses, account IDs, or hostnames are published. The dashboard itself carries the live detail; this page keeps to the story.