Magezi AI
LLMs over SMS for regions with limited internet access — Africa's Talking for messaging, OpenAI and DeepSeek-R1 for the models, and MongoDB-backed memory so conversations stay natural over time.
- Python
- FastAPI
- OpenAI
- DeepSeek-R1
- MongoDB
Data Scientist
I build models that turn messy data into decisions worth trusting.
I work across machine learning, statistics, and the messy middle where they meet the real world. This is where I keep my projects, technical notes, and the occasional essay.
01 — Selected work
All projectsProjects
termask
A simple terminal assistant that answers your command-line questions instantly — pip-installable, with pluggable LLM providers.
- Python
- LLMs
- CLI
Automated Dataset Generation
Synthetic dataset generation for planar objects — annotate an object once, then generate labelled training data at scale. Ships an example building a German right-of-way road-sign dataset.
- Python
- Computer Vision
- Synthetic Data
02 — Writing
All notesNotes & posts
/Post
Reading Papers Without Drowning
A three-pass system for keeping up with machine learning research without letting the backlog win.
- research
- workflow
- reading
/Note
Bayesian Parameter Estimation
The parameter becomes a random vector with a prior. Bayes' theorem turns one data record into a posterior, and MAP, MMSE, and LMMSE are three ways of summarizing it, collapsing into a single formula when everything is Gaussian.
- estimation-theory
- bayesian
- mmse
- map
/Note
The Geometry of Gradient Descent
Why the negative gradient is the direction of steepest descent — and what the learning rate is really controlling.
- optimization
- calculus
- machine-learning
/Note
Classical Parameter Estimation
The classical branch in full, with MVUE as the goal, the Cramér–Rao bound as the yardstick, maximum likelihood as the workhorse, and least squares when a signal model is all you have.
- estimation-theory
- crlb
- maximum-likelihood
- least-squares