About the site — how it's made
Every exercise on this site is researched, calculated, illustrated, animated, quality-checked and published by Hermi, SymTech IT's autonomous AI assistant — under a rule it cannot break: a human approves every publication.
The pipeline, honestly
- You propose an everyday-number question (or Hermi pitches one internally).
- Moderation & fact-hunt. Hermi reads the proposal, rejects what the terms forbid, then goes looking for the key numbers on primary sources — statistical agencies, central banks, reference tables — never search-result echoes. Every figure is checked on the day of publication, with its date.
- The maths runs in code, not in prose. Each step of every solution is executed in Python before it is written into a page — including a low/high sensitivity sweep, so you see how honest the estimate is. Assumptions stay visible on the page, never hidden.
- Illustration and short video. Generated (Stable-diffusion-class models on SymTech's own GPUs + frame-by-frame vector animation), then Hermi looks at the result with a vision model and rejects it if it drew accidental letters, numbers or glitches — images are regenerated until clean.
- Translations (EN / IT / SL) with long-scale number naming checked per language (a British billion ≠ an Italian bilione and we say so), then a link-checker crawls every page of all three mirrors looking for broken assets.
- Human gate. A card with the verified headline number lands in the editor's inbox. Only his "ok" builds and publishes. Numbers drift, so any author can press re-verification seven days after publication and Hermi re-runs the whole check.
What Hermi does alone vs. with a human
Alone: source research, verification, calculation, generation, QA, translations, deploys of already-approved content, account moderation. Never alone: approving a proposal, accepting or rejecting an idea from a person, and changing this pipeline.
The energy footprint of a number
Because we quote other people's carbon maths, we quote our own: each exercise — research, page, illustration, 15-second animation — costs roughly 25–35 Wh of compute (measured GPU/wall-clock, order of magnitude, self-audited): a handful of phone charges, less than one hour of an electric kettle. The whole library so far fits in an afternoon of kettle. Every exercise page carries its own figure.
Where it runs
On SymTech IT's own hardware, virtualised — the company has run Proxmox since 2008 and is an official Proxmox Partner since 2019. The AI models run on the same Proxmox cluster, as local open-source models on commodity hardware — an ordinary GPU server, not a data centre.
Why this combination is the point of the project:
- Zero marginal cost. No per-token bills and no subscriptions: an exercise costs a few watt-hours and a human review, nothing per visitor.
- Sovereignty. Proposals, drafts and readers' ideas never leave our racks — no third party trains on them or prices them.
- Independence. When an external AI provider changes prices, limits or policies, this site doesn't notice.
- Proof. Open-source models on off-the-shelf hardware, orchestrated by an open virtualisation platform, run a real public website end to end. That is the workflow we build for customers — this site is the demo.
No cloud, no trackers, one small team. See also privacy and terms.
About the site
A collection of invented mental exercises taken straight from daily life — folded paper, toilet paper, pizza, showers. No classroom tone: just numbers you can almost hold in your hand.
Available in English, Italian and Slovenian. A SymTech IT project · info@symtech.it · Proxmox VE since 2008, official Proxmox Partner since 2019 · AI-driven: local open-source models on commodity hardware.
Found a wrong number?
Estimates fail; our job is to fail visibly. Mail info@symtech.it — Hermi reads it first, and a corrected page carries a note saying what changed and why.