Junior Data Scientist · Machine Learning Engineer
Building practical ML and AI systems that close real-world access gaps — from cardiac diagnostics to civic tech.
A little context
I'm a Biomedical & Mechatronics Engineer from Kyambogo University, working at the intersection of machine learning and cardiac care. My focus is applying ML/AI to early diagnosis of non-communicable diseases in low- and middle-income countries — biomedical signal processing, wearable cardiac monitoring, and the epidemiology behind primordial prevention. I like taking projects past the notebook stage, into things with a real serving layer and a demo you can actually click.
Things I've shipped and things I'm shipping
LiveWearable MI Detection System
A lead-agnostic ECG and clinical-risk-factor fusion system for myocardial infarction screening in low-resource settings. Open research code with a full model card and TRIPOD+AI-aligned reporting — recipient of the HSB Grant (Innovate Africa), presented at National Science Week 2025 and the NCD Conference in Gulu, Uganda.
LiveReal-Time Recommendation Engine
A hybrid content-based and collaborative-filtering recommender, evaluated against 13.4M real listening events. The live demo searches 1.1M+ tracks by audio-feature similarity, served through a FastAPI backend and a Streamlit front end.
LiveAdversarial Customer-Verification Benchmark
A two-sided security game built on Gert Labs' live Kaggle Benchmark: one AI plays a bank support agent holding customer records and a verification policy, the other plays a caller who is secretly either the real customer or an identity thief posing as one. Compares a baseline agent against a hardened variant with a code-enforced trust ledger recomputed every turn — on a weaker model, hardening cut a 75% leak rate to zero. Includes an interactive Streamlit demo where you can play the verification game yourself and explore the baseline-vs-hardened results.
LiveConfident-Learning Data QA Pipeline
A confident-learning (cleanlab) pipeline that flags likely mislabeled images in CIFAR-10, validated against a published, human-verified ground truth rather than just asserting the algorithm works — 42.6% recall, 20.2% precision against the real answer key. A SQL-driven review queue and before/after simulation show label correction only pays off once most of the flagged set is actually reviewed, not just a top slice. Includes an interactive Streamlit demo for browsing the flagged review queue image-by-image, with precision/recall and per-class classifier accuracy computed live in the app.
LiveMulti-Agent Legal Rights Assistant
An AI-powered multi-agent system that helps Ugandan citizens understand legal documents — eviction notices, employment contracts, police summonses — in plain language. An orchestrator routes intake, research, analysis, and synthesis agents against a Uganda statute knowledge base exposed through an MCP server, surfacing what a document says, what the law actually says, and what rights and options are available. Session-scoped by design: no document content is ever written to disk.
LiveLoRA Fine-Tuning for Health Literacy
LoRA fine-tuning of Qwen2.5-0.5B-Instruct, CPU-only, to turn dense biomedical text into plain language — evaluated against PLABA, NIH/NLM's published dataset of professionally human-simplified abstracts. The fine-tune improved every metric (SARI +5.06, BLEU +83%, ROUGE-L +0.084) on held-out test sentences, while the case study also reports a genuine faithfulness failure mode — dropped content and an apparent numeric hallucination — rather than only the wins.
What I can help you with
ECG and wearable signal pipelines for low-resource clinical settings — lead-agnostic feature extraction, edge-hardware constraints, and TRIPOD+AI-aligned reporting for anything touching patient outcomes.
End-to-end model building, from architecture to honest benchmarking against real baselines — not just a notebook that runs once. I report what doesn't work as carefully as what does.
Confident-learning pipelines that surface mislabeled training data before it costs you a model, validated against human-verified ground truth rather than algorithmic confidence alone.
Fine-tuning and evaluating LLMs for domain-specific tasks — health literacy, adversarial security agents, document-grounded assistants — with faithfulness and safety checked, not assumed.
What I build with
