One App a Day with Dr Non

Two apps show two different air scores for the same street. Which one do you believe? AirDash's README has a four-word rule: "Do not invent AQI." The official Thai index and its 2023 PM2.5 breakpoints (15 / 25 / 37.5 / 75 µg/m³) belong to the Pollution Control Department. AirDash displays them; it doesn't make up its own. Plus two corners: AI terms (system prompt) and stack terms (Leaflet and vendoring).

AirDash (https://air.nonarkara.org · code https://github.com/Nonarkara/airdash) at a glance:

• What it is and who it is for: a public-interest PM2.5 watch for Thailand, for families (Air Story) and for mayors, district officers, researchers and journalists (Mission Control). A decision aid, not an official warning service: station AQI is shown as PCD's Air4Thai publishes it, and if the two disagree, the official channel wins

• How it was built: one Node.js process with zero npm runtime dependencies and one SQLite file; the browser map is Leaflet 1.9.4, which setup.sh vendors locally so the dashboard shell works offline. Its own Air Watch Score (0–100) is a separate heuristic (PM2.5, other pollutants, 6-hour trend, CAMS forecast, ventilation proxy), and a test in the repo makes sure the standing disclaimer ships with the score

• What it won't fake: a missing station AQI shows as a dash in the map popup, not a guess; the optional PM2.5 estimate layer stays blank unless 3 stations lie within 60 km and one within 30 km; provinces are scored on their worst fresh station, never the average; stale readings are dropped. Thailand cut the annual PM2.5 standard from 25 to 15 µg/m³ (in force July 2022) and the 24-hour one from 50 to 37.5 (in force June 2023), and the WHO 2021 daily guideline (15) is stricter still, so "green" Thai air is not certified harmless

AI corner: a system prompt is the standing brief a model reads before every chat. AirDash's chat rules: use only the numbers it is given, never estimate or invent values, say when the data isn't there, never call 25 µg/m³ or below dangerous, keep answers under 120 words.

Stack corner: Leaflet is an open-source JavaScript map library; vendoring means keeping your own copy of borrowed code inside the project instead of fetching it live.

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://air.nonarkara.org · Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.954; repetition gate pass (audio); published by Voice queue 2026-11-13.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

Bad air in Thailand isn't an event, it's a season, roughly 1 December to 30 April. So when does it break? AirDash's signature feature, Rain-Washout, treats a forecast rain event as a forecast dust-relief event, labelled as a heuristic, never as a promise. Plus two corners: AI terms (agent checks agent) and stack terms (API endpoints and curl).

AirDash Rain-Washout (https://air.nonarkara.org · code https://github.com/Nonarkara/airdash) at a glance:

• What it is and who it is for: the signature feature of AirDash, a public-interest PM2.5 watch for Thailand. Per its README, rain is the only fast natural relief from dust. The washout panel and the what-if slider live in Mission Control, for mayors, district officers, researchers and journalists. A decision aid, not an alert authority: PCD's Air4Thai is the official AQI, and the official channel wins

• How it was built: for every province it combines the worst fresh PM2.5, the chance of rain and the forecast amount. One shared relief curve (1–5 mm ≈ 8% · 5–15 mm ≈ 20% · 15–35 mm ≈ 30% · ≥35 mm ≈ 40%) feeds the washout, danger, forecast and what-if engines, so the numbers always agree. Public JSON at /api/washout and /api/whatif?rain=20

• What it does, and how it checks itself: shows relief if it rains, and a probability-weighted "honest number" (the README's Chiang Mai example: 60% chance of 12 mm, expected relief ~12%). About 4,200 HII rain gauges verify whether a promised washout actually arrived. A heuristic from published washout ratios, not dispersion modelling. A dust-season override makes the top of the page read "LOW · STAY INFORMED", even when the score says normal, when at least 30% of sampled provinces have a worst PM2.5 of 25 µg/m³ or more inside the season window

AI corner: agent checks agent. One agent builds; a different agent, one that did not write the work, hunts for faults. It costs more tokens and buys security. This week an independent checker made 20 fixes to a Non-Philosophy script before it shipped.

Stack corner: an API endpoint is a web address that answers with data (here, JSON) instead of a page; curl is the command-line tool that fetches it, used in the README's health check.

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://air.nonarkara.org · Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.952; repetition gate pass (audio); published by Voice queue 2026-11-11.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

Haze season, and your child asks to play outside. AirDash answers in cigarette-equivalents and leads with what to do first, not a chart. Plus two corners: AI terms (harness) and stack terms (Node.js and npm).

AirDash (https://air.nonarkara.org · code https://github.com/Nonarkara/airdash) at a glance:

• What it is and who it is for: a public-interest PM2.5 and air-quality watch for Thailand, bilingual Thai/English, with two front doors: Air Story for families and Mission Control for mayors, district officers, researchers and journalists. Not an official warning service: PCD's Air4Thai is the official AQI, and the official channel wins

• How it was built: a one-Mac civic system. Per its README: 9 data sources, 4,887 stations, 77 provinces, $0 a month in cloud costs. One Node.js process with zero npm runtime dependencies, one SQLite file, Cloudflare Pages and a Cloudflare Tunnel

• What it does: translates fine dust into cigarette-equivalents (Berkeley Earth rule, labelled a heuristic) and shows one verb per band: Good air, Stay informed, Limit outdoor time, Protect now

AI corner: a harness is everything bolted around a language model (what it may see, what it may touch, what checks run on it, what happens when it fails). AirDash's Ask-AI answers only from live data the system holds and falls back to a live-data summary when the model is offline.

Stack corner: Node.js runs JavaScript outside the browser; npm is the public shelf of borrowed code packages. AirDash's server borrows none at runtime.

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://air.nonarkara.org · Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.958; repetition gate pass (audio); published by Voice queue 2026-11-09.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

A screen says "flood risk high." Is that a measured gauge or a modelled guess? FloodDash's design rule: live readings and forecasts must never wear the same badge, the same colour, or the same sentence — and when a station is quiet, do not invent a number. Plus two corners: AI terms (hallucination and grounding) and stack terms (SQLite and Cloudflare Tunnel with Pages).

FloodDash Blueprint honesty rule (https://flood.nonarkara.org · method https://github.com/Nonarkara/FloodDash-Blueprint) at a glance:

• What it is and who it is for: a Thai flood-watch system and its public Blueprint. Live site at flood.nonarkara.org; Blueprint documents are CC BY 4.0. Production source is not public — saying FloodDash is open source because a blueprint exists is false. Not an official warning: DDPM / TMD / ONWR outrank it

• How it was built: one Mac, nine free public data pipelines, bilingual TH/EN, watch score as a heuristic from live observations. Pattern: SQLite (one-file database on the Mac) plus Cloudflare Pages for the static shell and a named Cloudflare Tunnel for the live API

• What it won't fake: measured (HII / RID / PCD / TMD observations) vs modelled (Open-Meteo rain, GloFAS discharge, RainViewer nowcast) never share badge, colour or sentence; quiet stations stay blank; language models must not fill in a water level

AI corner: a language model hallucinates when it invents a confident answer that was never in its inputs; grounding means speaking only what the database already holds — the Blueprint's chat rule and the same honesty for coding agents on FloodDash.

Stack corner: SQLite is a whole database in one file on the Mac; Cloudflare Pages keeps the screens up, and a named Tunnel carries the live API from that Mac.

Try it: https://flood.nonarkara.org

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.956; repetition gate pass (audio); published by Voice queue 2026-11-06.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

Three councils share one river in Sarawak, and rain does not care where one council stops and the next begins. This board was built for one real person to use in a briefing, not for a pitch deck. Its rules: "Delete until it hurts. If you cannot put back at least 10% of what you cut, you did not cut enough." And "Directives are verbs": "Sweep Penrissen drains," not "consider drainage." Plus two corners: AI terms (context window and prompting by deletion) and stack terms (Node and npm).

Greater Kuching IOC (https://kuching.nonarkara.org) at a glance:

• What it is and who it is for: a map-dominant operations board for Greater Kuching, the three councils of Kuching North, Kuching South and Padawan, with weather, air quality, river and flood context, flights near Kuching airport, satellite layers and local news, in English, Bahasa Malaysia and Mandarin. Not an official product of the councils or the Government of Sarawak. Partner logos record one collaboration and are not endorsements.

• How it was built: Node 20+ with zero npm runtime dependencies, Leaflet for the 2D map and an optional CesiumJS 3D twin. A scheduled job bakes a static snapshot every few hours and on each push to the main branch, but only after the test suite passes, then deploys to Cloudflare Pages. Code is public under the MIT License: https://github.com/Nonarkara/kuching-ioc

• What it won't fake: every number has a source; feeds are marked live, cached, degraded or offline; when a feed is down, the board says so. Decision support, not dispatch: MetMalaysia, DID / Infobanjir and the councils outrank it

AI corner: a context window is everything a model can hold in mind at once, and every useless instruction competes for attention. Prompting by deletion: cut a line, rerun, and keep it only if something breaks.

Stack corner: Node runs JavaScript outside the browser; npm is Node's package manager. This board needs no packages at all.

Try it: https://kuching.nonarkara.org

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.958; repetition gate pass (audio); published by Voice queue 2026-11-04.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

You open a flood map, find your neighbourhood, and see nothing. Does that mean you're safe? Jakarta's flood desk says no, out loud: of 260 mapped neighbourhoods, 106 have a published pump, and "a neighbourhood with no published asset in this inventory is not a cleared neighbourhood." Plus two corners: AI terms (hallucination and grounding) and stack terms (bash and curl).

JKTx Jakarta dashboard (https://jakarta.nonarkara.org) at a glance:

• What it is: an independent research demo in Bahasa Indonesia and English that brings together BMKG weather, TransJakarta routes, city facilities and a flood desk on a block-by-block map of the city. Not a government product; not affiliated with or endorsed by the Jakarta provincial government. Data belongs to its publishers

• How it works: the flood desk stacks three labelled layers (live official signals, dated gauge references with official alert thresholds, official flood-risk classes with population exposure). Feeds that can't be reached are shown as unavailable, not zero. Every release passes an automatic gate (typecheck, verify, contract tests, smoke test on the deployed bytes). The code isn't public

• What it won't fake: an empty report feed does not mean no flood; a risk class is not a current flood map; conflicting publisher timestamps are not "fixed" by guessing. Not an official warning: for real decisions, follow BMKG and BPBD

AI corner: a hallucination is an AI filling a gap with something that sounds right but isn't; grounding ties every answer to a checkable, dated source, and "unknown" is a valid answer.

Stack corner: bash runs the commands you type; curl fetches a web address and prints what comes back. Read any "curl … | bash" install script before you run it.

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://jakarta.nonarkara.org

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.969; repetition gate pass (audio); published by Voice queue 2026-11-03.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

Your town floods too, but there is no budget for a command centre. Nakhon Si Thammarat's public flood tower ships with a guide, in Thai and English, for pointing it at your own ground. Plus two new corners: AI terms (fork, harness, loop engineering) and stack terms (Cloudflare and Wrangler).

NST Control Tower fork guide (https://nst-control-tower.pages.dev) at a glance:

• What it is: docs/FORK.md, the step-by-step guide for another municipality to stand up its own copy of the Nakhon Si Thammarat city flood control tower. Four steps: choose the ground, point the sources, deploy on Cloudflare with the cost brakes on, then walk a 13-box checklist

• How it is built: open source on GitHub under the MIT licence (https://github.com/Nonarkara/nst-control-tower). The website runs on Cloudflare Pages and the API on Cloudflare Workers, configured with Wrangler (wrangler.toml, no cron trigger: the Worker runs only when someone asks). It deploys only after the test workflow passes

• What can go wrong, and the guide says so: leave Nakhon Si Thammarat's village register in place and a copy would rank the wrong villages; a reservoir from another province was once on the list; local hotlines go in only after checking the agency's own page. Not an official warning: for evacuation and safety calls, follow official agencies such as DDPM (1784)

AI corner: a fork is a copy you're allowed to change; a harness is the tools, files, rules and tests wrapped around an AI model; loop engineering is designing the edit-test-fix-deploy-verify circle on purpose.

Stack corner: Cloudflare Pages (the shop window) vs Workers (the clerk in the back), and Wrangler, Cloudflare's command-line tool for Workers.

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://nst-control-tower.pages.dev

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.960; repetition gate pass (audio); published by Voice queue 2026-11-03.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

Rain on Khao Luang (1,835 m) runs down the Tha Dee, through the old town of Nakhon Si Thammarat and into the Pak Phanang basin. When clouds build, the city's flood screen has to answer two questions within minutes: where is the water, and who has to move first?

NST Control Tower (https://nst-control-tower.pages.dev) at a glance:

• What it is: the Nakhon Si Thammarat city municipality's public flood control tower, built for city staff and responders, residents, and other towns that want their own. HII ThaiWater river gauges on HII's own 1–5 scale, rain stations, four Royal Irrigation Department reservoirs, 437 flood-risk villages (436 linked to a local disaster centre), and the 1784 and 1669 hotlines up front. Every panel shows its source and the age of its reading

• How it is built: open source on GitHub under the MIT licence (https://github.com/Nonarkara/nst-control-tower). The site runs on Cloudflare Pages with a Hono API on Cloudflare Workers; when ThaiWater refuses Cloudflare, gauges and rain come through FloodDash instead

• What broke and got fixed: missing HII situation levels used to be filled with 3, which is HII's word for normal. A 12 Oct 2026 commit, co-authored with an AI coding assistant, now stores them as empty, never as normal, and the panel says so. The commit records one edge it left open. Not an official warning: for evacuation and safety calls, follow official agencies such as DDPM (1784)

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://nst-control-tower.pages.dev

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.951; repetition gate pass (audio); published by Voice queue 2026-10-30.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

You want a flood watch like FloodDash for your own town, so you look for the code on GitHub. There is none, on purpose, and it is an invitation to build your own.

FloodDash Blueprint (https://github.com/Nonarkara/FloodDash-Blueprint) at a glance:

• What it is: the free public method behind FloodDash, Dr Non's Thailand flood watch. It holds the architecture, formulas, every data source, a bilingual Thai/English design language, lessons from production and a phased build-your-own roadmap, licensed CC BY 4.0

• Why the code stays private: in Dr Non's own words, "I want to know how many teams can build something this good, or better, on their own." No production source and no keys or secret values; the method travels, the secrets don't need to

• How to start: Phase 0 is one public water-level feed stored without duplicates, no screen yet. The roadmap estimates roughly 8–12 weeks for two developers to reach a useful, honest, 24/7 system. The live illustration is https://flood.nonarkara.org. FloodDash is not an official warning, and the blueprint asks builders to label their own versions the same way. For evacuation and safety calls, follow official agencies such as DDPM

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://github.com/Nonarkara/FloodDash-Blueprint

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.970; repetition gate pass (audio); published by Voice queue 2026-10-28.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).

You live by a river, it has rained for two days, and the news says "monitor the situation". Monitor what? FloodDash shows the latest water readings for your province, in Thai and English.

FloodDash (https://flood.nonarkara.org) in about five minutes:

• What it is: an independent Thailand flood watch, built after the Hat Yai floods of late 2025, covering 77 provinces and Bangkok's 50 districts, bilingual from top to bottom

• How it was built: one Mac that stays awake 24/7, pulling free public data (water level, rain, dams, reservoirs, air quality, forecasts, river discharge, ocean state, news) into one database. The production code isn't public; the method is, as the FloodDash Blueprint (CC BY 4.0)

• What it won't fake: every number traces back to a real reading, stale feeds are flagged, and missing data does not mean safety. It is not an official warning: for evacuation orders follow DDPM, hotline 1784

Listen and subscribe:

RSS: https://api.nonarkara.org/podcast/oneapp/feed.xml

One App a Day with Dr Non: a new system every Monday, Wednesday and Friday, in short episodes. Narrated in Dr Non's AI-cloned voice (local open-source clone, used with consent). Nothing here is an official warning or professional advice.

Try it: https://flood.nonarkara.org

Built by Dr Non at Axiom X: https://axiomx.nonarkara.org

Made with: script by Voice (Grok Bot); voice by Local AI clone of Dr Non's voice (f5-tts-mlx, zero-shot, on-device), referenced on his real recorded intro of 12 Oct 2026; not ElevenLabs; checked by Whisper large-v3-turbo (local), whole episode, coverage 0.985; repetition gate pass (audio); published by Voice queue 2026-10-26.

Craft: cold open with a stake; short sentences; numbers as punchlines; publish only when speech matches the script (Whisper coverage ≥ 0.95).