JengaAI

Open-source ML training and inference framework for African AI.

Train multi-task NLP, Speech, and Vision models with a single YAML config — no code required. Built in Kenya, built for Africa.

What We Build

JengaAI is a framework that lets researchers, engineers, and non-technical teams train production-grade machine learning models on African language data — and deploy them without vendor lock-in, without API dependencies, and without sending sensitive data to foreign servers.

Your model. Your data. Your task.

Models

Model	Task	Language	Base
Rogendo/afribert-kenya-adapted	Masked Language Modeling (DAPT)	Swahili · Sheng · English	castorini/afriberta_large
Rogendo/cpims-nlp-intent-urgency	Intent + Urgency Classification	Swahili · Sheng · English	afribert-kenya-adapted

afribert-kenya-adapted

Domain-adaptive pre-training of AfriBERT on ~39M tokens of Kenyan language data — Swahili Wikipedia, East African journalism, synthetic Sheng/code-switch corpus, and real CPIMS field worker WhatsApp data. Achieves 30.4% average perplexity improvement over the base model on Kenyan domain text, with 66% improvement on Sheng and 41% on English-Swahili code-switching.

cpims-nlp-intent-urgency

Multi-task classifier trained on CPIMS child protection support messages. Simultaneously predicts 63 intent classes and urgency level (high / medium / low) from a single encoder pass. Intent F1: 74.5% — up from 46% on a generic English base model. Handles English, Swahili, and Kenyan code-switching.

With this framework, the possibilites of languange and Natral language processeng are limitless!

The Framework

pip install jenga-ai

Train any model with a single YAML config:

project_name: swahili-hate-speech

model:
  base_model: castorini/afriberta_large
  max_seq_len: 128

tasks:
  - name: classification
    type: single_label_classification
    data_path: data/hate_speech.csv
    text_column: text
    label_column: label

training:
  epochs: 5
  batch_size: 16
  learning_rate: 3.0e-5

python -m jenga_ai train --config swahili-hate-speech.yaml

Supported modalities

Modality	Status	Notes
NLP — classification, NER, multi-task	✅ Production	Multi-task with shared encoder + dual heads
Speech — Whisper fine-tuning, transcription	⚙️ Active development	ASR for Swahili and African languages
Vision — classification, OCR, object detection	⚙️ Active development	Document verification, image classification
LLM — LoRA fine-tuning, Ollama integration	⚙️ Active development	Swahili instruction tuning

Key capabilities

Multi-task learning — one encoder, multiple task heads, shared representations
Domain adaptation — continued MLM pre-training for African language domains
Responsible AI built in — explainability engine, audit trail, human-in-the-loop routing, bias evaluation
Offline-first — trained models run without internet, no per-query API cost
HuggingFace native — load any HF model as base, push trained models to Hub
No-code web platform — upload CSV, click Train, get predictions

Why JengaAI Exists

Africa's AI ecosystem is being built on API wrappers — products that call GPT-4 or Claude and rebrand the output as "African AI." These products are expensive at scale, dependent on foreign infrastructure, unable to handle African languages properly, and unable to keep sensitive data on the continent.

JengaAI exists to make the alternative practical.

A locally trained, domain-adapted model:

Costs nothing at inference time after training
Runs fully offline in low-connectivity environments
Can be fine-tuned on your specific institutional language
Keeps sensitive data — health records, case notes, financial transactions — under your control
Performs significantly better on African languages than generic multilingual models

Use Cases

Child protection systems — intent classification and urgency triage for CPIMS support messages in English, Swahili, and Sheng

Community health — symptom extraction and referral urgency from CHW voice notes and field reports

Financial services — M-PESA dispute classification, fraud signal detection, transaction intent analysis

Government services — citizen complaint routing, document OCR, service request classification

Education — student question routing, learner sentiment analysis, multilingual content classification

Media monitoring — hate speech detection, misinformation flagging, topic classification in Swahili and code-switched text

Responsible AI

JengaAI is built with responsible AI development as a core principle, not an afterthought:

Explainability — every prediction can be explained in human-readable terms
Audit trails — hash-chained tamper-evident logging of all inference decisions
Human-in-the-loop — low-confidence predictions are flagged for human review
Bias evaluation — evaluation across language, demographic, and domain subgroups
Data sovereignty — models run locally, data never leaves your infrastructure
Transparent limitations — model cards document what the model gets wrong, not just what it gets right

Community

JengaAI is developed in the spirit of African AI communities doing the work right — Data Science Africa, Masakhane, Deep Learning Indaba, and AIMS.

We believe that building AI for Africa means building it on African data, in African languages, with African institutional contexts — not wrapping foreign models in local branding.