DrukView — Every Ngultrum Accounted For

Your dzongkhag. Your data. Your right. An open-source civic-transparency platform for Bhutan: where public money goes — development projects, budgets and allocations by dzongkhag (and gewog/village, where documents go that deep) — with the funding source and a link to the original document on every figure.

What you get

• Home: dzongkhag search + an interactive map of Bhutan (real boundaries, bundled with the app so it always renders): zoom (+ / −, Ctrl + scroll, pinch), drag when zoomed, hover names. Click a dzongkhag → it zooms in, shows its 205-gewog drill-down (click a gewog to filter its budget lines) and opens the budget explorer (fiscal-year selector, clickable funding-source / sector / gewog bars, expandable projects).
• Per-dzongkhag dashboard (/en/thimphu) with a module sidebar, and a national one (/en/bhutan): Overview · Budget & allocations · Projects · Funding sources · Weather & rainfall · Tenders · Leadership & offices.
• Compare dzongkhags, Data sources (live status per source), Update log, CSV export, "Report a problem" links, a "last refreshed" stamp on every module.

Stack

Next.js 15 (App Router, TypeScript) · PostgreSQL (pg, plain SQL migrations) · Vercel Cron collectors · Vitest.

Data sources

Source	How	Refresh
World Bank Projects & Operations API	automated (pipeline/worldbank.mjs)	daily
Asian Development Bank IATI file	automated (pipeline/adb.mjs; override with ADB_IATI_URL)	daily
Parliament of Bhutan member records (National Assembly + National Council)	automated (pipeline/leaders.mjs; override with PARLIAMENT_API)	daily
Kuensel and The Bhutanese feeds — leadership-change watch	automated (pipeline/news.mjs; override with KUENSEL_FEED / BHUTANESE_FEED)	daily, with the members
e-GP — Electronic Government Procurement System (PMDD, Ministry of Finance) — every tender currently out to bid (203 notices on 2026-10-11)	automated (pipeline/egp.mjs; override with EGP_BASE)	daily
Open-Meteo weather	automated (pipeline/weather.mjs)	daily (Vercel Hobby allows at most one run/day per cron; on a Pro plan change vercel.json to e.g. 37 */6 * * *)
Ministry of Finance, National Budget Report FY 2026-27 — national summary, sector allocation, and the allocation of every dzongkhag, thromde and all 205 gewogs	reference file data/reference/budget-2026-27.json	yearly (new budget each May)
NSB 2017 Population & Housing Census (Table 2.1)	reference file data/reference/population-2017.json	per census
Ministry of Health, Annual Health Bulletin 2025 (Table 6.1) — all hospitals, PHCs, sub-posts, HISCs with beds	reference file data/reference/health-facilities.json	yearly
MoESD Annual Education Statistics 2024 (Annexure IV) — all 566 schools by name and level	reference file data/reference/schools-2024.json	yearly
Parliament member records, as shipped	reference file data/reference/leaders.json — the fallback shown until the daily refresh above has run	on change
MoICE Annual Industry Report FY 2025-26 (Annexure 3) and National Employment Report 2024-25 — industry licences and jobs filled per dzongkhag	reference file data/reference/economy.json	yearly
Dzongkhag administration websites and dated appointment notices — Dzongdags and office contacts	reference file data/reference/dzongdags.json, each entry graded by how well it is confirmed	on change
MoF / MoHCA / MoAL & other ministries' budgets, allocations, tenders, offices	curated CSV in data/manual/ — these are PDFs/web pages with no API. Every row must cite its public document	on commit

Reference files ship with the code and work even without a database. tests/reference.test.ts checks each against the publication's own totals (census sums, the Health Bulletin's per-dzongkhag totals rows, MoESD Table A, 47 NA seats with the constitutional split), so a transcription slip fails CI.

Collectors run via vercel.json crons → /api/cron/<job> (rate-limited per job; a CRON_SECRET bearer skips the limit), and from the Load / refresh data now button on the Data sources page. Each job is isolated: on failure the previous good rows stay, the source is flagged stale on the Data sources page, and the run is written to the update log. A batch with any invalid record loads nothing.

When a collector stops

A job that runs and fails flags its own source stale. A job that stops running at all is the harder case: the status stays ok and the data simply ages, which nothing inside the app notices. So /api/cron/alerts checks the age of each source's last success once a day against a 72 h limit (STALE_ALERT_HOURS), and an overdue source is reported in three places:

• on the Data sources page, so a reader can see the figures are that old;
• to ALERT_WEBHOOK_URL, if set — one POST, at most once a day per source, plus a line when it recovers. The body carries the message under text, content and message, which covers Slack, Google Chat, Discord and ntfy without a per-service adapter;
• at /api/health, which answers 503 while any source is overdue (degraded: "stale-sources") or the database is unreachable (degraded: "database").

Point any uptime monitor at /api/health: it is the only one of the three that still works when the scheduler running the other two has itself stopped.

How the leaders list stays current

The daily leaders job reads Parliament's own member records and replaces the list — but only if the response is a complete Parliament: 47 National Assembly seats with the constitutional split per dzongkhag, and one elected National Council member per dzongkhag. Anything less is rejected, the previous list stays, and the source is marked stale, so a half-written or broken response can never leave a gap on the site. The job finds the current term itself (the highest-numbered Parliament that is fully seated), so it keeps working after the next election without a code change. Until the job has run, the bundled data/reference/leaders.json is shown instead.

Newspapers are a watch, never a source. The same job reads the Kuensel and The Bhutanese feeds and flags articles that read like a change of office-holder in a named dzongkhag. A flag is a dated link inviting someone to re-check that entry against the official source — no name on DrukView is ever taken from a newspaper, because a headline such as "Dasho takes charge in Trongsa" cannot be parsed into a correct name, and a wrong name on a transparency site is worse than a late one. A feed that fails never fails the members refresh.

How the tender list is collected

Bhutan's e-GP portal publishes no API, feed or export, and its tender page is a shell: the rows are fetched by the page's own script. The daily tenders job makes that same request — POST /TenderDetailsServlet with funName=AllTenders, carrying a session cookie from a prior GET of the search page — and reads the bare run of <tr> rows it returns. This request was observed against the live portal, not inferred; the first version of this collector guessed at it, fetched nothing for a day, and is written up in ADR-0002.

What it collects is the portal's live register: viewType=Live is 11 pages of 20, 203 notices when last measured, every one still open. The job reads all of it — 11 requests, about 13 seconds — so DrukView holds every tender out to bid rather than an arbitrary slice. The portal's other views are not read: the unfiltered view is 731 pages of history, Archive is 685 pages of closed notices and Cancelled is 37. A page size above 20 makes the servlet drop the connection, so the page size stays at 20.

Because that response's shape is the whole agreement, the parser is the tested part (tests/egp.test.ts runs it against a captured real response) and the job is all-or-nothing: a response without a readable row fails the run, the previous notices stay published and the source is marked stale. A redesign at e-GP therefore shows up as a stale source, never as missing or invented notices — and the Tenders page says so in as many words, because "no tenders here" and "we could not read e-GP" are different claims and only one of them is about Bhutan.

A tenders page shows at most 500 notices, one more than that is fetched so a cut-off list can say it is cut off rather than passing for the whole register, and the e-GP live register held 203 when last measured. Three honesty constraints: e-GP publishes no estimated value, so DrukView shows none; each notice cites its own page on e-GP (ViewNotice.jsp?id=…, which opens without a session) rather than a search page; and a notice is placed in a dzongkhag only when the procuring agency's own hierarchy names exactly one — never from a place name in the title, where it is usually where the work is rather than who is buying it. Of 203 notices, 129 were placed in a dzongkhag and 74 left national.

Contributing## Contributing

DrukView is open source (MIT) and open to anyone. Start with CONTRIBUTING.md — it sets out the seven data rules (never invent a number; cite every row; never mix Nu. and US$; never guess an identity or a place; reject the batch, not the row; say what you do not know; keep the date), how to add cited CSV rows or a reference dataset, and how to write a collector.

The most valuable contribution needs no code: if a figure here is wrong, open a Data correction issue with a link to the public document that proves it. Also see the Code of Conduct and, for anything security-related, SECURITY.md — report privately, never in a public issue. CI runs typecheck, tests and a production build on Node 20 and 22 for every push, plus a check that every curated row cites an https document and that no secret is committed.

Honesty rules (enforced in code and tests)

• No sample/estimated data ships. With no data, the UI says so.
• Nu. (government) and US$ (partners) are never summed or converted.
• A multi-dzongkhag project is counted in each dzongkhag but its money is not split — only single-dzongkhag projects add to a dzongkhag's US$ total.
• Locations guessed from place names are labelled "inferred from text". "Committed" is never shown as "spent".
• Weather is modelled for each dzongkhag's headquarters town (approximate point), not station-measured.

Run locally

cp .env.example .env.local          # set DATABASE_URL (CRON_SECRET optional)
npm install
npm run db:migrate && npm run db:seed   # optional: the app also does this itself on first visit
npm run job -- manual               # load data/manual/*.csv
npm run job -- worldbank            # needs network access to the source
npm run dev                         # http://localhost:3000

Tests: npm test (unit). Database integration tests run when TEST_DATABASE_URL points at a throwaway Postgres database (CI does this automatically; the suite drops and recreates the public schema there).

Deploy (Vercel)

1. Import the repo into Vercel and attach a Postgres database (Neon, via the project's Storage tab). That sets DATABASE_URL — the only required setting.
2. Open the site. The tables are created automatically on the first visit.
3. Go to Data sources → Load / refresh data now (about a minute). Every source then also refreshes automatically once a day via Vercel Cron.

Optional: ALERT_WEBHOOK_URL (where to send a notice when a source has not refreshed for three days — any Slack/Google Chat/Discord/ntfy incoming-webhook URL) and STALE_ALERT_HOURS to move that limit.

Optional: CRON_SECRET. The data jobs only read public data, so they run without a secret but at most once per interval (World Bank/ADB and Parliament 6 h, e-GP tenders 3 h, weather 2 h, curated CSVs 10 min; 5 min after a failure). A request carrying Authorization: Bearer $CRON_SECRET skips that interval. Every job stops after 50 s and records why, so a slow source can never fail silently.

Data-source verification (2026-10-10)

The automated sources were checked against their real responses, and the parsers' tests use trimmed copies of those (tests/fixtures/*real*):

• World Bank: 63 Bhutan projects. Handles project_abstract = {"cdata!": …}, US-style dates (12/31/2027 12:00:00 AM), and totalcommamt "0" (= not reported, shown as such, never "US$0").
• ADB (IATI): https://www.adb.org/iati/iati-activities-bt.xml is valid IATI 2.03; 62 activities are merged into 38 projects (e.g. grants 0088 + 0089 → project 39221-022, US$14.30M, matching adb.org). IATI status 3 "Finalisation" = completed (adb.org shows these as Closed). Sector codes are resolved to names from the official IATI codelist (shared/iati-sectors.json).
• Map boundaries are bundled (src/geo/bhutan-dzongkhags.json, public/geo/bhutan-gewogs.json) from OCHA ROAP / HDX via geoBoundaries (gbHumanitarian 2019, CC BY 3.0 IGO): 20 dzongkhags + 205 gewogs from one consistent dataset; every gewog's dzongkhag was assigned geometrically and checked against the official count per dzongkhag (tested).
• Open-Meteo: multi-location response format matches the parser.

What is still unproven is the runtime on your host (network egress, timeouts, database permissions): run the collectors once after deploying and read the Data sources page.

Known gaps

• Tenders: only e-GP's live register is read, so a notice that closes drops out of DrukView at the next refresh rather than becoming an archive — the portal's Archive (685 pages), Cancelled (37) and Contract-award views are not read yet, and reading them would mean a different, much larger job. e-GP publishes no estimated value, so the Value column is empty for its notices. A notice whose procuring agency is a ministry is national, even where the work is plainly in one dzongkhag.
• World Bank records carry no sub-national location, so most are "national / unspecified"; ~10% are placed by place names in the text (labelled "inferred").
• Dzongkha: the app is wired for locales (/en/…) but ships English only; a Dzongkha translation will be added once a native speaker can review it.
• Population is the 2017 census (the next census is due 2027); per-person ratios that divide newer counts by 2017 population are labelled indicative.
• Health: Trongsa's totals row in the Bulletin reports 4 sub-posts but names only 3.
• Schools: names come from PDF text extraction (a few spellings may be off); the middle- vs higher-secondary split per dzongkhag is approximate (combined figure is exact).
• Dzongdags: there is no single official machine-readable list, so each name is read from the dzongkhag's own site or a dated appointment notice and is published with a confidence grade (4 confirmed by a dated notice, 14 from an undated website page, 2 unconfirmed). Re-check them against RCSC appointment notifications when they can be reached.
• MoICE publishes no unemployment rate by dzongkhag; its labour-force figures come from the NSB Labour Force Survey and are national only. Industry figures are licences in force, not businesses known to be trading.
• Village-level allocations depend on curated rows; the Ministry of Finance budget goes down to gewog level only.
• 31 of the 205 gewog names in the budget could not be matched to a map boundary with confidence, so those gewogs are listed in the tables but left unshaded on the map rather than attached to the wrong shape.