# ⚚ I reverse-engineered my mobile operator's APK — then Hermes Agent wrote the executive report > Source: > Published: 2026-05-28 23:01:45+00:00 *This is a submission for the Hermes Agent Challenge: Write About Hermes Agent* // 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" } Just check how amazing these reports look like! "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**. 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. 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. 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: 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. **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:** One message: *"update with fresh data"* — triggered 8+ tool calls automatically: DuckDB → diff → scripts → charts → LaTeX → 2× compile → verify. 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. 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. **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. ** Qwen3.7-Max via OpenRouter.** Alibaba's flagship agentic model, built for long-horizon autonomous execution. **326 requests. 61.4M tokens. $19.57. ~1 hour. vs a full week manually.** The next step: **run it all locally**. I'm currently eyeing an [Apple Mac Studio M4 Max](https://www.ldlc.com/fiche/PB00675416.html) — 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. With 64GB unified memory, I'd be able to run: The same class of model as Qwen3.7-Max on OpenRouter — locally, for free, forever. A fully local Hermes Agent stack. *From $19.57 on OpenRouter to owning the hardware. That's the roadmap. 🇳🇨* The pattern: **monitor → query → visualize → compose → translate**. The *translate* step — turning a percentile into a story a CEO can act on — is where Hermes Agent earns its keep. *Have you used Hermes Agent for multi-tool orchestration? Curious how your experience compares.*