AI in Research

Mapping a whole university catalog. 1,780 unique modules in the AY2025/26–AY2026/27 catalog. 85 pass the sustainability screen (27 focused / 58 inclusive); 34 pass the entrepreneurship screen (12 / 22).

Only 1 module lands in both sets — BF3122 Carbon Markets — and 0 make the strong GreenComp-Acting × EntreComp-Into-action join.

Phrase coverage inside the 85: GreenComp 28 · greening topics 62 · SDG 7 of 85. Empty fields are outline-language gaps, not proof a topic is untaught.

Method: screen → tag → crosswalk → audit. Whole-word matching, false-friend stripping and community-college prefix overrides. Domain words that decide inclusion stay on the screen and are never copied into the 12 competences — that would double-count the same phrase.

AASHE / EntreComp inclusion rule: title hit, or two distinct terms in title + outline. UNESCO ESD 8 derived from GreenComp via the Wiek 2021 crosswalk — not a second screen.

A national training catalog, checked against demand. 25,813 courses in the official directory; 2,137 match the green / ESG screen (8.3%); 122 official SWDA green skills; 2,789 course–skill links written.

Coverage of the 122 official skills: 29 with no matching course · 7 thin (1–2 courses) · 17 moderate (3–8) · 69 crowded (9+) · 0 validation against live pages.

Where supply is missing, by sector: Sea Transport 5 · Landscape 5 · Energy and Power 4 · Food Manufacturing 3 · Carbon Services 3 · Financial Services 2.

Next move: build sector-technical modules against the empty rows, and pair disclosure reporting with carbon and circular-operations courses so a report can be evidenced.

Demand is an official list. Supply is catalog text. A gap here is a catalog gap — enrolment and job outcomes are not in the open files.

An RNA-seq pipeline that runs on a laptop. Counts in → QC + CPM filter → DESeq2 → GO + GSEA → report.

• Not FASTQ — takes an iDEP-style gene × sample count matrix, so it runs without a cluster.
• Hard stops — fewer than 3 replicates per group, or batch confounded with condition, halts differential expression.
• Two engines — DESeq2 with apeglm shrinkage; limma-trend if the matrix is FPKM/TPM.
• Browser mode — a Cloudflare Worker re-runs case-vs-control in PyDESeq2; files never leave the page.

Demo run: 68 genes up · 1 gene down · 1,998 genes after CPM filter · 31.2% variance on PC1. Worked example: 6 samples, 2,000 genes, padj < 0.05 and |log2FC| ≥ 1. Top gene RNU6-163P at log2FC 2.65, padj 6.1e-11.

PCA separates case from control before differential expression is trusted; Hallmark GSEA runs on the full ranking with no hard cutoff; summary.json + HTML report carry full sessionInfo. GO enrichment returned nothing significant at this demo size — reported as-is rather than dressed up.

Six keyless datasets become ecosystem accounts:

• io-lulc v2 — 10 m annual land cover, 2017–2023
• CTrees AGB — 100 m annual biomass, 2000–2025
• Sentinel-1 RTC — 10 m SAR backscatter, VV and VH
• MODIS 13Q1 — 250 m NDVI, 16-day composite
• MODIS 64A1 — 500 m monthly burned area
• geoBoundaries — ADM0–ADM3 units, 48 Asian countries

Output shape — SEEA-EA accounts: extent (hectares by class) · condition (NDVI, SAR, burned) · services (carbon stock, CO₂e, 2000–2025 series) · historical change — each with a provenance block and an uncertainty flag beside the number.

Method trap: uncapped counts every pixel (7.5 s per cell; Singapore 13.2M pixels; fails past ~44M px) and is correct for area. max_size decimates to 2,048 px (0.8 s per cell; Singapore 2.2M pixels) and holds shares but under-counts area by 6×. The rule written into the engine: uncapped for area, capped for shares — and every scaled number is flagged estimated.

Before and after, in the room’s own language. The pre-course form asks for a one-sentence problem statement, 1–10 familiarity with AI tools, 1–10 practice on own work, root-cause clarity / confidence / feasibility, and what has already started. The post-course form repeats the same 1–10 scales, requires a Yes/No on whether an AI tool was used today, and asks what changed in the proposal.

Bilingual English + Chinese throughout; anonymous; email not collected.

Instrument design: no direct motivation item (it ceilings at 8–10; infer motivation from actions already taken) · familiarity ≠ practice · anchor every item to the person’s proposal · drop items that cannot move in one class · ship every item, scale label and choice bilingually, not just the question stems.

Limit: a single class cannot show whether the two-week pitch improved. The class scales measure the class; a third round at the pitch is what would measure the outcome.

From a 77-minute recording to filed cases. Voice transcript (two recordings, ~77 and ~18 min) → AI draft (minutes + case write-ups from the transcript) → case records (EL-01 to EL-07, one per practice) → human correction (names, dates and titles fixed before filing).

The date came from the recording timestamps, not from anything said in the room — and the AI sub-cluster’s driver was left open for the presenters rather than invented.

A proposal where AI drafts but never invents the evidence.

• Phase 0 · 8–12 weeks — waste-split analysis, screens fractions, returns three candidate products. No airport machine.
• Phase 1 · on campus — Changi feedstock on the miniTRASHPRESSO. First article, test sheet, LCA on measured yield.
• Phase 2 · one site — a single product, stream and location, bought under green procurement. Separate decision.

Claims the draft was held to: no Miniwiz municipal carbon figure as a Changi result; no invented throughput, power or machine feedstock; FY2024/25 waste split not reassigned to FY2025/26.