Tsela & Co. · Investor MVP · 28 October 2026

Ads written in isiZulu, Swahili and Hausa, checked before they run

We trained an open, commercially licensed language model to write short Facebook ads in three African languages, tested it on 150 briefs per language, and pointed it at a real website.

3 languages450 test briefs5 training rounds~3 h on one H100 GPU (about R200)

Result

Pass rate on 50 locked tests per language

Locked tests are briefs the model never trained on. The model shown is round 4, chosen by its score on the practice tests, not the locked ones. "Before" is the same open model with no training.

Live example

We pasted tselaco.com and asked for three ads

The model read the site, planned three ads (a pain, a proof and an offer angle), then wrote each in the chosen language. Ad words are real text set over a generated background, never painted into the image. Under each ad is the English back-translation from a second, independent model.

Backgrounds: FLUX.1 schnell, checked automatically for people, stray lettering and blank frames. All three passed on the first try.

Training

Swahili clears the 85% bar, Hausa sits just under it, isiZulu needs more data

Each round trains a small add-on (LoRA) on licensed text plus the practice ads that passed, then reruns all 450 tests. Locked scores move by a few points round to round because each language has only 50 locked tests.

Method

How each ad is checked

Every ad must pass all seven checks. They are free, open-source and run on our own GPU, so checking costs nothing per ad.

Right language: GlotLID identifies the language with at least 60% confidence.
Meaning kept: Two different models translate the ad back to English; both must match the brief (LaBSE similarity).
Meaning, direct: The ad itself is compared with the brief across languages.
Real words: At most a quarter of words may be missing from a word list built from licensed text.
Business name: The name appears exactly as given.
No prices or numbers: Brand rule: prices never appear in ads.
Length: Headline up to 45 characters, text up to 160.
Translation benchmark: FLORES+ test sentences scored with chrF++ and AfriCOMET, used only to test, never to train.

Models and data, all licensed for commercial use

Writer: Google Gemma 4 E4B with our trained add-on (Apache-2.0).
Second checker: Cohere Aya-101 (Apache-2.0). Tested as a writer too; it scored 0% because it often answers in English or ignores length limits.
Training text: WURA web text (Apache-2.0) for all three languages, plus the practice ads that passed. No output from Claude, GPT or Gemini was used for training.
Dropped after round 1: a government isiZulu corpus whose text is lower-case with broken words. It taught the model bad habits.

Limits

What the machine checks still miss

All nine website ads passed the machine checks, yet reading the back-translations shows real mistakes. The pass rates above are an upper bound until a native speaker reviews a sample.

Next steps

Native-speaker review of 30 ads per language (about R2,000 to R3,200) to measure true accuracy and tune the checks.
More and cleaner isiZulu training text, then rerun the same 150 tests.
A grammar check for isiZulu word agreement (ZulMorph), and a stricter meaning check for calls to action.