---
name: thermograph-orientation
description: What Thermograph is as a product and a business, how the estate is shaped, and the principles behind how it is run.
---

# Thermograph, from the top

## The product

Thermograph answers one question for any point on Earth: **how unusual is the
weather right now, here?** It grades recent and forecast conditions against ~45
years of local climate history (1980–present, ERA5), turning raw numbers into a
percentile and a plain tier — Normal, Above/Below Normal, High/Low, Near Record.

The mechanic: a clicked or searched location snaps to a stable **~4 sq mi grid
cell**; that cell's full daily history is fetched once and cached. Each day is
placed on *its own* day-of-year ±7-day window as an empirical percentile. So
"97th percentile" means hotter than 97% of the last ~45 years of this same
fortnight, at this exact spot.

Precipitation is graded differently, because most days everywhere are dry:
a day with measurable rain is ranked only among the **rain days** in its window,
spreading wet days across intensity tiers. Dry days are coloured by dry streak
instead.

### The framing rule

Grade names are **relative to each location's own climate**, never absolute
temperature words. A 60°F day can be Above Normal in a cool coastal summer and
Below Normal in a hot inland one; the same day can be Near Record for January
and ordinary for July. Calling a percentile "hot" or "warm" implies a scale the
data does not carry. Every grade means "unusual **for this place, this time of
year**." Preserve that in any copy, summary or chart you produce.

## Who it is for

- **Weather-curious people** — "is this heat actually unusual here, or does it
  just feel that way?" The answer is a defensible percentile against that exact
  location's history, not a national average.
- **Movers, travellers, planners** — the city climate pages and compare view.
- **Alert subscribers** — told when their place crosses into genuinely unusual
  territory, without daily noise.

## What exists

Map and grade view, a two-year calendar heatmap, single-day breakdown, location
compare, ~750 SEO-indexed city climate pages, a PWA with an offline shell.
Accounts are optional and exist to power alerts: per-location subscriptions on
any metric with a percentile threshold, delivered by in-app bell, Web Push,
or Discord DM, with a weekly cap and per-event dedup so one alert is at most one
notification a week. A versioned public JSON API (`/api/v1`, `/api/v2`).

**What deliberately does not exist:** any monetisation — no plans, ads or paid
tiers — no native mobile apps, and no user analytics beyond anonymous counters.

## The estate

The live fleet is exactly three hosts on a WireGuard mesh — not two VPSes plus
a desktop, that generation of topology is gone:

| Host | Runs |
|---|---|
| **central1** (mesh `10.10.0.251`) | **testing** (testing.thermograph.org) as its own Swarm stack with its own TimescaleDB, OneDev (the forge, `code.jinemi.com`), Grafana + Loki (`/opt/observability`), Taiga (tasks.jinemi.com), Postfix for the whole fleet, and the NetBird server (human/device access to the mesh) |
| **prod1** (mesh `10.10.0.250`) | **prod** (thermograph.org) as its own Swarm stack, Centralis itself, nightly backups |
| **cibuilder** (mesh `10.10.0.252`) | OneDev's own build agents — not a Swarm node |

`prod1` and `central1` each manage their own independent, single-node Swarm —
never one shared swarm, and never a "central1 manager, prod1 worker" split.
NetBird (on central1) is for people's laptops and phones joining the estate's
network; it does not replace the WireGuard mesh above, which still owns
fleet-to-fleet (host-to-host) routing.

The two Contabo VPSes this used to run on, **vps1** and **vps2**, are fully
decommissioned — not frozen, not reachable for reads, gone outright, along
with the `dev` environment that lived on vps1 (dev.thermograph.org, retired
2026-08-31; `dev` the branch is unaffected). The operator's Mac holds the
estate's keys (SSH, the age key, OneDev/OpenBao tokens) but runs no
environment of its own and is not a fleet member; a phone reaches the estate
only as a terminal, for hands-off work via git.

`Jinemi/thermograph` is a monorepo (`backend/ frontend/ infra/ observability/`)
with per-domain CI. `Jinemi/docs` is separate — the planning layer, on its
own cadence.

## How it is run

A small team, plus provisioned automation agents that work through exactly the
same door as a person: branch → PR → required CI → explicit merge. The
guardrails live in the platform rather than in anyone's memory — which is what
makes it a team's estate rather than one person's, and what lets an agent be
given the same door without a second set of rules.

The principles, as actually practised:

1. **Own the whole stack.** Georedundant SaaS replaced by three netcup hosts
   and discipline. The whole estate costs less monthly than one managed database.
2. **Privacy as a feature.** No tracking cookies, no per-visitor IDs, allowlisted
   anonymous counters only. Cheap to keep, expensive to retrofit.
3. **Boring prod.** Anything clever proves itself on testing first; prod
   changes arrive only as already-proven image tags.
4. **Everything is a PR.** Person or agent, infra or app or docs. The route in
   is a PR against OneDev at `code.jinemi.com` — through Centralis's `onedev_*`
   tools from anywhere, or a plain `git push` plus the OneDev web UI from a
   shell; the door is the same either way.

## Where to look next

- `connectors` — what to install; OneDev's own `onedev_*` tools, which is how
  ordinary git work is done here, arrive with Centralis and need no install.
- `docs_search` — architecture decisions, runbooks, and the business layer
  (product overview, operating model, costs, roadmap).
- `notes_search` — what people wrote down along the way.
- `skill id="thermograph-ops"` — deploying and debugging.
- `skill id="thermograph-community"` — Discord.
- `skill id="thermograph-notes"` — where written work belongs.