SSIC 2025 Decision Trail

Describe a business (or paste a MyCareersFuture job link to load the employer) and watch TypeSafe Jev walk the Singapore Standard Industrial Classification 2025 tree, level by level, to a 5-digit industry code — animated as a pixel-art crowd on a 2D canvas, with every probability listed.

Live: https://ssic.danielquek.com · Sister app: SSOC 2024 Decision Trail

Run

npm install
npm start          # http://localhost:5173 — needs TYPESAFE_API_KEY in the environment

Optional env: PORT, TYPESAFE_MODEL (default jev-latest).

Deploy to Cloudflare (production: https://ssic.danielquek.com)

npm run deploy

worker/index.mjs serves the same API as server.mjs from a Worker; public/ is static assets. Set the TypeSafe key once as a Worker secret: npx wrangler secret put TYPESAFE_API_KEY.

Spend protection (in wrangler.jsonc): /api/classify is limited to 5 runs/min per IP and a global DAILY_LIMIT of 300 runs/day (KV counter USAGE, ~23k input tokens per run); /api/mcf to 30 requests/min per IP. Local test: CLOUDFLARE_INCLUDE_PROCESS_ENV=true npm run dev:cf.

How the decision is made (lib/classify.mjs)

1. Read — a Noul asks whether the text describes what a business does (asked alongside level 1).
2. Levels 1–5 — Section (A–V) → Division → Group → Class → Sub-class. Beam search (width 3): each level is one Jev request holding one Choice per path still in the beam; options are the node's direct children, described with SingStat definitions, the sub-class titles they contain, alternative titles from the alphabetical index, and "see …" exclusions. Paths are ranked by geometric-mean edge probability (TypeSafe hierarchical-classification cookbook).
3. Finals — the 5 best sub-classes meet in one Choice with full definitions, covered activities and exclusions. Its distribution is the final probability list.

Six requests in total, ~1.4 s. The server streams each step as NDJSON so the canvas animates live.

Reading the canvas

• Trail (top): READ → SECTION → DIVISION → GROUP → CLASS → SUB-CLASS → FINALS. Click a stop or use ←/→.
• Crowd: every option Jev weighed, sorted by tier — pink leading, orange kept in beam, gold considered then pruned, grey little chance. Each of the 82 divisions has its own costume (chef for food service, hard hat for construction, lab coat for R&D…). Hover for details.
• Top paths: segmented bars show path probability; bubbles show the chain (vote at this level × probability so far).
• Finals: podium heights = final probabilities; result card compares with the employer's own SSIC tag on MyCareersFuture.
• Decision log below the canvas lists every question and probability as tables.

Files

Path	What
server.mjs	HTTP server, /api/classify (stream), /api/mcf (load employer from a posting)
worker/index.mjs	Cloudflare Worker with the same API plus rate limits and a daily cap
lib/classify.mjs	Beam search + finals using @typesafe-ai/sdk
lib/mcf.mjs	MyCareersFuture client: employer profile + SSIC tag
public/	Canvas app (app.js), procedural pixel sprites (sprites.js), division costumes (costumes.js), sample companies (samples.json)
data/ssic2025.json	Built from SingStat xlsx in raw/ by npm run build:data
tools/snapshot_samples.mjs	Refreshes public/samples.json from samples/mcf_samples.txt
tools/eval.mjs	Runs the sample companies, compares with employer tags
tools/check_costumes.mjs	Every division has a distinct, valid costume
TEST_RESULTS.md	Latest evaluation

Data: SingStat SSIC 2025 detailed definitions, alphabetical index and SSIC 2020↔2025 correspondence tables.