Built for how Asia actually writes.
We are building privacy tools for people who mix languages in one sentence. Here is what works today, where the gaps are, and how we will measure our progress.
Language is only the first layer.
Working across ASEAN means tone, formality, names and formats change from one market to the next. We design for three layers.
Language
Scripts, vocabulary and mixed-language writing: Malay, Chinese, Thai, Vietnamese, Indonesian, Khmer, Lao, Burmese and English, often in one message.
Today: transcription and translation in 20 languagesCulture
Formality, honorifics, politeness and sentiment that differ by country, so a reply or a classification fits the audience.
Custom work: sentiment, classification, conversational AIContext
Local formats: names, addresses, dates, currencies and national ID numbers, which decide whether personal data is found or missed.
Planned: MY, SG, TH, ID, VN identifiersWhere Apple’s on-device models stop, we start.
Our on-device apps build on the language tools Apple ships in iOS. They cover major languages well, but not the ones below, which millions of people in ASEAN use every day.
Translation and transcription we support today
- English
- Mandarin
- Cantonese
- Japanese
- Korean
- Hindi
- Arabic
- Thai
- Vietnamese
- Indonesian
- Spanish
- French
- German
- Italian
- Portuguese
- Russian
- Dutch
- Polish
- Turkish
- Ukrainian
Gaps we want to close
- Malay (Bahasa Melayu)
- Khmer
- Lao
- Burmese
- Tagalog
Why “Manglish” breaks most privacy tools.
People in Malaysia, Singapore and across ASEAN blend English with Malay, Chinese, Tamil, Thai or Vietnamese within a single message. Detectors trained on one language at a time miss names, IDs and addresses that appear mid-sentence, in another script, or in an unexpected format.
A tool that promises to “protect personal data” must handle this, or it gives a false sense of safety. So we treat code-switched text as a test case of its own, and will publish where we fail.
Boss, pls email the contract to siti.aminah@example.com lah, her IC is 900101-14-1234, and she stays at No. 12, Jalan Contoh ya.
Illustrative sentence with fictitious details. The address, the email and an IC number in a Malaysian format appear in one Manglish message. Today, AirGap PDF AI’s rules would catch the email and probably the address, but not the IC number.
Regional ID formats we plan to detect
| Country | ID | Format | Why it is tricky | Our apps today |
|---|---|---|---|---|
| Malaysia | NRIC / MyKad | 12 digits: YYMMDD-PB-###G | a date-of-birth prefix means false positives are likely without context | Not yet |
| Singapore | NRIC / FIN | Letter + 7 digits + check letter (e.g. S1234567D) | the check letter allows validation | Not yet |
| Thailand | National ID | 13 digits with a check digit | often written with spaces or dashes | Not yet |
| Indonesia | NIK | 16 digits | easily confused with other long numbers | Not yet |
| Vietnam | CCCD (citizen ID) | 12 digits | older 9-digit CMND cards still appear in documents | Not yet |
How we will measure, and why there are no scores yet.
We have not run these tests, so we are not publishing numbers. When we do, this table will show our results beside a cloud model and Apple’s own tools on the same test sets, including where we lose.
| Test | Languages | Metric | Test data | Result |
|---|---|---|---|---|
| Personal-data detection | English, Malay, Mandarin, Thai, Indonesian, Vietnamese | Precision and recall per entity type | Labelled documents in each language | Planned |
| Code-switched text | Manglish, Singlish, Bahasa Rojak, Chinese–English, Thai–English | Precision and recall; failure cases published | Real-style messages with fictitious personal data | Planned |
| Regional ID formats | MY NRIC, SG NRIC/FIN, TH ID, ID NIK, VN CCCD | Detection and false-positive rate | Synthetic and format-valid test IDs | Planned |
| Transcription | Languages supported by Apple Speech | Word error rate | Short recorded samples, including accented speech | Planned |
| Policy analysis | English, Malay, Chinese | Agreement with a lawyer’s review of the same policies | Real public privacy policies | Planned |