Lionel
Ebebe.

Problem

Housing search in Cameroon runs on WhatsApp threads and cash deposits, which makes listing fraud cheap and payment disputes near-impossible to resolve. There was no trustworthy booking primitive.

Approach

A React Native / Expo client over a core API, with BullMQ workers and a WebSocket layer for live booking state, and a separate FastAPI risk service scoring listings. Fraud detection combines logistic regression and random forest classifiers with perceptual hashing to catch re-posted listing photos. Payments go through MTN and Orange Mobile Money aggregators, reconciled by both webhooks and a poller, with 10-minute booking holds and a double-entry ledger so every cent has two sides.

Outcome

Reached 200+ users. The interesting engineering was the money path: idempotent webhook handling, a poller as the reconciliation backstop, and a ledger that makes disputes auditable instead of anecdotal.

Problem

Mortgage files arrive as scans of scans. Analysts needed to ask questions across a thousand-document corpus without hand-reading it, and needed to know when the system was wrong.

Approach

Built a Python OCR ingestion pipeline, chunked and embedded the corpus with MiniLM, indexed with FAISS, and served answers through Mistral-7B behind a Gradio UI. Evaluation was the point: I built a labelled query set and tracked Recall@5 and MRR across chunking and retrieval configurations rather than eyeballing outputs.

Outcome

Recall@5 of 82–87% and MRR of 0.74–0.82 on the evaluation set, with the retrieval configuration chosen from measurements instead of intuition.

Problem

A growing student organisation was running on scattered social posts and a static page nobody could update.

Approach

React + TypeScript on Firebase, shipped across several development cycles: a Reels-style media feed, an events system with full create/read/update/delete for organisers, an E-board section, authenticated admin access, and scroll-driven motion throughout.

Outcome

A maintained platform the org's board actually updates itself, iterated across multiple semesters rather than shipped once and abandoned.

Problem

Framed around a Johannesburg urban-heat-island scenario (SDG 13): simulating heat diffusion across a city grid large enough that a single process is the wrong tool.

Approach

Implemented the simulation in MPI and compared 1-D against 2-D domain decomposition, and blocking against non-blocking halo exchange, measuring scaling behaviour and communication overhead across process counts.

Outcome

A benchmarked account of where communication cost overtakes compute — and why 2-D decomposition wins as the process count grows.