Selected Work/Fig. 02 — systems, held in tension
Systems built to survive contact with real users, real data, and real failure.
EnBW Energie Baden-Württemberg AG (via Ishango.ai)·Data Platform Engineer → MLOps Engineer
Keeping production ML and data workflows alive across multiple teams at a large European energy utility — deployment, observability, and working out why a pipeline died at 3am.
First author
A vendor-agnostic architecture for agentic incident triage, diagnosis and guarded remediation across data and AI pipelines — assembled from open-source parts rather than bought.
Lumis·Founder
Finds why a data or ML system broke — from its own alert to the change that caused it — and tests every explanation against evidence it gathered read-only. Across 117 investigations with four models on the GridCast estate, 11 of 15 injected faults were never missed.
Noeud·Founding Machine Learning Engineer (R&D)
A probabilistic FX intelligence backend for Ghanaian importers and exporters: 30-day forecasts with block-bootstrap uncertainty, and an LLM layer that weighs them against market news and bank rates while model baselines stay immutable and audited.
MinoHealth AI Labs·Machine Learning Engineer
An autonomous tool-using research agent and the distributed platform under it — 20+ scientific tools behind one API, multi-day runs, 20,000+ candidates per batch. It produced four publications.
Agents in production, computer vision in the field, and independent research.
Tokenisation, attention, the training loop — from first principles, so the thing I operate all day isn't a black box to me.
A deep research agent that plans, retrieves, reaches for tools and synthesises — without a human nudging it at every step.
The public, safety-bounded deployment of the research agent — same tool infrastructure, without needing access to any of it.
Detection of crop stress and disease from the air, plus the drone capture and mapping pipelines that fed it. Flown over real farms in the Bono and Savannah regions.
An end-to-end pipeline testing whether gold-market session structure is predictable at all. Mostly an exercise in not fooling yourself with a flattering backtest.
Prior art. Where the assumptions above were first tested.