I’m Brian Kerio, a ML Engineer who loves coding, and solving real-world problems with technology. I studied Computer Science and Mathematics at Kenya Methodist University. I’m fascinated by what’s possible when engineering meets bold thinking. Forever learning, building, and solving.
Tech Stack: TensorFlow | Keras | Streamlit | OpenCV | Google Gemini API
🔗 Repo: https://github.com/BKerio/Briantumor-classification
- Classified brain MRI scans into glioma, meningioma, pituitary tumor, or no-tumor using a transfer-learning Xception model (97.56% test accuracy) and a custom CNN trained from scratch (91.16%).
- Built an interactive Streamlit app with gradient-based saliency maps and Gemini-powered AI explanations of model focus areas.
- Trained on ~7,000 Kaggle MRI images with a stratified train/validation/test split.
Tech Stack: Python | scikit-learn | Pandas | NumPy | Matplotlib
🔗 Repo: https://github.com/BKerio/Diabetes-prediction
- Built a production-style ML pipeline on the Pima Indians Diabetes dataset, comparing Random Forest, Gradient Boosting, and SVM classifiers.
- Applied 5-fold cross-validation and hyperparameter tuning, evaluating with F1-score, ROC-AUC, and Matthews Correlation Coefficient.
- Automated missing-value and outlier handling in a modular, reusable preprocessing pipeline.
Tech Stack: Python (standard library only) | Custom autograd engine
🔗 Repo: https://github.com/BKerio/SquirrelGPT
- Implemented a GPT-style transformer from scratch in under 100 lines, with zero ML framework dependencies — no PyTorch, TensorFlow, or NumPy.
- Hand-built a scalar-level autograd engine, causal self-attention, and character-level tokenizer to expose the mechanics behind LLMs.
- Deliberately overfits a small dataset as a teaching tool for backpropagation, embeddings, and attention.
Tech Stack: ESP32 | Node.js | Express | TypeScript | MongoDB | Daraja API
🔗 Repo: https://github.com/BKerio/PoolPay-Automation
- Built an IoT access-control terminal: users enter a phone number and amount on a keypad, guided by an LCD, to pay via M-Pesa STK Push.
- ESP32 polls payment status every 3 seconds and triggers a servo-driven gate unlock on confirmed payment.
- Hardened the backend with Helmet, rate limiting, and env-based credential validation; documented extensibility for a Stripe gateway alongside M-Pesa.
Tech Stack: WhatsApp Cloud API | Node.js | TypeScript | Express | MongoDB | Socket.IO | Daraja API
🔗 Repo: https://github.com/BKerio/whatsapp-bot
- Built a conversational WhatsApp assistant for Millenium Solutions handling service inquiries, user registration, and a team directory.
- Integrated M-Pesa STK Push payments via the Daraja API, pushing real-time transaction status to clients over Socket.IO.
- Delivered through the Meta Cloud API with a TypeScript, path-aliased architecture.
