Ede / AI / NLP
A language problem, before a model problem.
A public African-language NLP experiment built to understand where general-purpose AI falls short.
01 / The tension
A fluent-looking answer can conceal uneven language performance. African languages are underrepresented in available language technology, and choosing a larger model does not automatically resolve that gap.
I tested language inputs, reviewed open datasets, and explored model and prompting approaches. Data coverage and evaluation emerged as constraints on what the product could reliably claim—not just implementation details to resolve later.
02 / The decisions
Choose a learning loop over a training stack.
I scoped Ede as a focused, public experiment using available models and data. The immediate job was to expose language behaviour and learn from it.
Large-model fine-tuning offered a higher potential quality ceiling, but required data and compute beyond the scope of an early experiment.
Treat evaluation as part of the product.
I made data exploration and observed language behaviour central to iteration. A change in model or prompt only mattered if there was a useful way to judge the result.
A simple AI wrapper could have shipped faster. It would also have left the underlying coverage and evaluation problem largely untouched.
03 / What I built
I owned the problem framing, scope, data exploration, model experimentation, and public release. Making the experiment accessible moved evaluation beyond my development environment.
04 / What changed
experiment released
A practical environment for exploring African-language behaviour with available AI approaches.
a defining constraint
The work identified data availability and evaluation quality as priorities for meaningful improvement.
The lesson I’m taking forward.
This is an experiment, rather than a demonstrated language-quality breakthrough. Its value is in making the limitations concrete. I would define the evaluation signal earlier, so that iteration could distinguish a convincing answer from a useful improvement.