Farid Dihan Nahdi

UK home-care agencies are inspected against Care Quality Commission standards. The engine reads a provider’s records and rates them the way an inspector would, but it throws out any finding whose quoted evidence is not in the records, and it votes across repeated runs so one lucky run cannot pass.

62% → 9%

fabricated findings, measured against a human auditor

2×

agreement with the human auditor (8/25 → 16/25)

~627

medication visits saved from a wrong “compliant” flag before release

For teams putting AI in front of regulators, auditors or clinicians.

Purchasing, budgets, double-entry accounting and project tracking in one system, replacing chat threads and spreadsheets. Supplier quotes are compared automatically, and over-budget payments or bank-detail changes need a second person’s approval.

34

staff in 6 divisions use it every day

23,356

material price records from 157 suppliers

17,727×

currency error found and fixed (US dollars stored as rupiah)

For companies whose money still moves through WhatsApp and Excel.

Engineers lay out panels on a map, simulate every hour of the year, and get a payback case with PLN export rules and local-content (TKDN) requirements applied automatically, instead of paying for a desktop licence per engineer.

11

customer pre-orders

8,760

hours simulated per site, per year

0.01°

sun-position match to the pvlib reference on 7 sites

For energy companies that sell on the strength of a feasibility study.

trendJack drafts social posts from each day’s industry news. No new customer could finish onboarding, and most workspaces had never published. I moved payment ahead of setup, let solo users approve their own drafts, and capped AI spend per workspace.

9 of 11

production workspaces unblocked to publish

~$0.22

AI cost per news sweep, capped at 5 per workspace a day

0 → 441

tests actually running, now on every pull request

For SaaS teams whose product works in a demo but not for paying customers.

Teachers in ordinary schools rarely have a special-needs expert nearby, so their questions waited for days. The assistant answers inside the WhatsApp groups they already use, moves sensitive questions to a private chat, and refers every case to a human expert.

142

teacher questions answered

27

teachers across 2 district governments

11.4 s

median reply time

For programmes that need to reach people where they already talk.

Three shops ran on separate systems and four databases, so every product was entered again and again. One platform now serves all three, each with its own look and domain, and a product entered once appears in every shop.

10,344

products migrated automatically

4 → 1

databases

2

legacy shop systems retired

For groups running several brands on duplicated systems.