01
Research
We study how open models can be evaluated and adapted for Nigerian and African contexts. Our work values reproducible experiments, clear limitations, and responsible publication.
Research, evaluation, and AI products for African contexts
Independent research and product lab
Reliable AI begins with evidence.
Nicom AI Labs develops research assets, evaluation methods, model adaptations, and domain-specific AI products for Nigerian and African contexts.
Our work starts with domain expertise, careful evaluation, and measured results.
We combine research, evaluation, and product development to build AI systems that are useful, responsible, and suited to real local contexts.
02
Evaluation
We build evaluation processes that test correctness, usefulness, safety, and domain fit. Expert review is central to this work.
03
Products
We translate validated research into focused AI products and partnerships. Product direction is informed by evidence, responsible design, and real user needs.
Current model release
NicomEdu-Alpha
NicomEdu-Alpha is a public experimental model specialized through fine-tuning for SS2 Mathematics classroom-style responses.
It is the first public NicomEdu model release from Nicom AI Labs. It is intended for experimentation and research. Verify important answers independently.
Status: Public Alpha model release
View NicomEdu-Alpha on Hugging Face
Hosted demo: coming soon.
Alpha limitations: The model can make mistakes, repeat text, or use inconsistent formatting.
Latest research release
NicomEdu Nigeria v0.1
NicomEdu Nigeria v0.1 is our multilingual education adapter — a LoRA fine-tune that translates in-depth English tutoring answers into Hausa, Igbo, and Yoruba while preserving the complete teaching sequence: definitions, logical steps, calculations, cautions, stated uncertainty, and the closing learner check-questions.
We built it on a clean open base, trained it on rights-cleared English↔Hausa/Igbo/Yoruba parallel data, and evaluated it on a locked, human-scored grid spanning a STEM core across all three languages, reviewed by qualified native speakers. Where the unadapted base model collapses into degenerate repetition, the adapter reproduces the full pedagogical sequence in-language and handles safety and exam-integrity refusals correctly.
Status: Non-commercial research release (CC BY-NC 4.0)
View the adapter on Hugging Face
Limitations: A translation instrument, not an original in-language tutor. It carries known STEM vocabulary drift, and all output is draft material for qualified human review — not for high-stakes use such as examinations, grading, or medical/legal/financial advice.
Operating principles
Careful by design.
- Ground work in domain expertise and local context.
- Separate private evaluation assets from public communication.
- Protect personal data and minimize collection.
- State limitations clearly.
- Publish selected methods and findings when evidence supports them.
Contact
Research and partnership enquiries
Contact Nicom AI Labs at raysamy20@gmail.com.