InitLabel logoInitLabel

Case studies

Selected projects from agritech, health, finance and AI labs. Client names are shared with permission or anonymised at the client's request.

AI-generated agricultural and computer vision visual
412k

crop leaf images labelled for 23 disease classes in 9 weeks

East African agritech startup · Image · Polygons · High accuracy

Challenge

A crop-disease app for smallholder maize and cassava farmers needed segmentation data from phone photos taken in poor field light. Existing offshore vendors confused maize lethal necrosis with nutrient deficiency.

What we did

Trained a 40-person team with agronomists from a Kenyan agricultural university. Built a reference library of 600 hard cases and added an agronomist review stage for low-confidence items.

Result

Gold-set accuracy of 98.9%. The client's model F1 on field data rose from 0.71 to 0.88, and the app now serves farmers in Kenya, Uganda and Tanzania.

“They knew what a sick maize leaf looks like in Kakamega in April. That's not something you can write into guidelines.”
Head of ML, client company
AI-generated speech and language data visual
3,800 h

of code-switched Swahili–English call audio transcribed

Regional telecom and mobile money provider · Audio · Transcription · Enterprise

Challenge

A customer-care voice assistant was failing on Sheng and mid-sentence switches between Swahili and English. The calls contained personal and financial data covered by the Kenya Data Protection Act.

What we did

Ran the project on an isolated secure floor with no internet access from workstations. Annotators transcribed and redacted national ID numbers, phone numbers and M-Pesa codes in the same pass.

Result

Word error rate on the client's ASR model fell by 34%. The ODPC-aligned audit found zero personal-data leakage incidents.

AI-generated LLM evaluation visual
96.4%

reviewer agreement on 180,000 LLM preference comparisons

US-based AI research lab · LLM · RLHF ranking · Dedicated team

Challenge

The lab needed preference data and red-teaming in Swahili, Amharic, Yoruba and Hausa to improve model safety in African languages, where public evaluation data is scarce.

What we did

Assembled a 60-person team of university-educated native speakers across Nairobi and partner hubs in Addis Ababa and Lagos, with weekly calibration sessions against the lab's researchers.

Result

Delivered two weeks ahead of schedule. The lab extended the engagement to a 12-month dedicated team.

AI-generated document processing visual
11 days

from kickoff to a production-ready dataset of 60,000 receipts

Pan-African fintech lender · Text · NER · Pay as you go

Challenge

A lender needed to extract merchant, amount, date and KRA PIN from handwritten and thermal-printed receipts to speed up SME loan applications.

What we did

Started with a 1,000-item pilot, then scaled through the API with 48-hour rush turnaround and webhook delivery straight into the client's training pipeline.

Result

Field-level accuracy of 99.2%. Manual document review time for loan officers dropped by 70%.