// Detect dark theme
var iframe = document.getElementById('tweet-2059698327235805258-252');
if (document.body.className.includes('dark-theme')) {
iframe.src = "https://platform.twitter.com/embed/Tweet.html?id=2059698327235805258&theme=dark"
}

🍿 Demo

📸 Screenshots

Just check how amazing these reports look like!

Building a Multi-Stakeholder API Monitoring Report with Hermes Agent

"Just generate a PDF" — famous last words. What started as a simple request turned into something much bigger: a full monitoring stack for my mobile operator, built evenings and weekends, culminating in a 17-page professional report that would have taken a full week to build manually. Total cost with Hermes Agent: $19.57.

🏝️ The Backstory: A Stack No One Else Has

I'm Adrien, a developer in New Caledonia. OPT-NC's Helia mobile service has an app — but no public API, no CLI, nothing on any marketplace. So I reverse-engineered the APK, extracted the private HTTP calls, and rebuilt them in a Go CLI that snapshots voice, data, and SMS consumption every 5 minutes into a local DuckDB database.

Then I built KDE Plasma widgets in Python/PyQt that read from DuckDB and display live on my desktop — mirroring the official app's data, but with history, trends, burn rate, and alerts. Plus a system tray icon showing live API status at a glance.

The official app shows you now. My stack shows you now, history, trends, and alerts.

No other Helia customer has this. That's the breakthrough.

🎯 The Problem Hermes Agent Solved

All this data was sitting in DuckDB. The question was: how do I present it to people who don't speak SQL?

Three completely different audiences: the CEO who needs a 30-second summary screenshot-ready for PowerPoint, the CIO who wants ROI in euros and SLA compliance, and the Network Admin who needs actionable tickets with specific hours and error patterns.

🎭 The Role-Playing Game That Became the Design Document

Before writing a single line of code, I asked Hermes Agent to do something unusual: simulate a full team meeting with 7 personas — CEO, CIO, network admin, developers, and marketing.

It produced a full transcript. Each persona argued their case:

• The CIO: "I need ROI in euros, not percentages."
  


• The CEO: "I need the 30-second version I can project tomorrow morning."
  


• The Network Admin: "Give me Jira tickets with specific hours, not a dashboard."
  

That transcript became the design document. Every page of the report was written against a specific person's stated need. No guessing. No generic output.

⚡ What Hermes Agent Built in ~1 Hour

It started with the data. No assumptions — it queried the schema and immediately caught something: average latency was 2,534ms but the median was 204ms. Bimodal distribution. That single insight shaped every chart.

It extracted brand colors from the website before writing a single line of LaTeX. Hermes Agent opened Helia's site, pulled the magenta/pink gradient from the SVG logo, and used it consistently across every chart, table, and tcolorbox. Small detail. Big difference.

Then it built everything:

• 4 Python scripts: latency distributions, timeout heatmap, sparkline, 4-panel executive chart


• Full XeLaTeX report with TikZ progress bars and brand-colored boxes


• A French accent fixer script baked into the pipeline

One message: "update with fresh data" — triggered 8+ tool calls automatically: DuckDB → diff → scripts → charts → LaTeX → 2× compile → verify.

📊 The Charts It Produced (Without Being Asked)

I asked for charts. I didn't ask for this:

Latency distribution — median/mean/P95 lines labeled, shaded fast vs slow zones, annotated arrow pointing to the long tail: "Queue longue (timeout ~12s) (~40% des pings)". Log-scale version revealing the true bimodal structure: two peaks at ~80ms and ~4s.

4-panel executive dashboard — availability gauge (87.2% vs 99.5% SLA), latency (2534ms vs 500ms — "5.1x trop lent"), timeout rate ("1 requête sur 8 échoue"), composite SLA score per metric. Score global: 65%. Verdict: RED.

Timeout heatmap — all 18 timeouts concentrated on Wednesday evening 19h–22h. The rest of the week: clean. Instant actionable insight for the network admin.

All in Helia brand colors. All annotated. None of it explicitly requested.

📋 The Executive Summary: Fits in One Slide

Page 3 of the report is designed to be screenshotted directly into PowerPoint. Verdict box at the top (red), plain-language translation ("1 call in 8 fails"), business impact in euros, top 3 problems with responsible parties. The CEO gets it in 30 seconds.

🧠 What I Learned

Orchestration is the real superpower. I asked for a PDF. It fetched brand colors, queried the DB, wrote Python scripts, compiled LaTeX — all unprompted.

Iteration is the default. The first charts were basic. The final ones have annotated thresholds, color-coded data points, and statistical summaries. That's the loop.

Skills are the real ROI. Everything got distilled into a reusable ~/.claude/skills/reporting-latex/ skill. Next similar project starts at 80%, not zero.

💰 The Cost

— 16-core CPU, 40-core GPU, 64 Go unified memory. New Caledonia is 20,000km from everything; local inference just makes sense. Zero API costs, zero latency to the cloud, full control.