Practice · Data
Models that run on the data you actually have.
Most of the data work that reaches us starts small and messy: a few thousand rows, a sensor that drops out overnight, a spreadsheet someone maintains by hand. We build for that rather than for the dataset we wish existed, and we teach it the same way: the architecture matters less than whether it survives contact with the data you can really get.
How the practice works
Teach
It runs as a taught track
Data science is one of the things OKHub teaches, not a service line bolted on afterwards. People arrive with a laptop and leave having trained a model, measured it honestly and put it somewhere it has to keep working.
Build
Then the same people build
The projects below were built the way they were taught: pick the smallest architecture that fits the signal, get it to run end to end, and only then make it faster.
Deploy
On the network that exists
Inference gets pushed to the edge, models get sized to run on the device in the room, and reporting assumes the connection drops. That constraint shapes the design from the start rather than being patched in at the end.
Work we have shipped
Each of these has a page of its own with the full build.
ANN + LSTM
Credit Card Fraud Detection
Sequence modelling over spending history, so a transaction is judged against the behaviour that came before it rather than in isolation. Exposed to banking systems over REST.
Read the build
GRU on raw audio
Speech Emotion Recognition
Feature extraction straight off the waveform, then a recurrent model over the sequence, because emotion is in how a sentence moves rather than in any single frame of it.
Read the build
YOLOv8 + CNN
Sign Language Educator
Detection to find the hands, classification to read the sign, both sized to run on an edge device so the feedback is immediate and nothing has to leave the room.
Read the build
Django + LLM
UTME Career Advisor
Scores and subject combinations matched against real university requirements, with the language model doing the explaining rather than the deciding.
Read the build
Sensor feeds + prediction
AI-Powered Agriculture
Soil and weather telemetry turned into a recommendation a farmer can act on that morning, delivered to a phone rather than a desktop dashboard.
Read the build
What we work with
- Sequence models
- LSTM, GRU
- Vision
- YOLOv8, CNNs
- Language
- LLM-backed flows behind Django
- Signals
- Audio features, IoT sensor telemetry
- Delivery
- REST APIs, dashboards, edge inference
- Where it runs
- Phones, edge devices, patchy connections