For founders & teams who want their AI to actually ship — multi-agent systems, RAG pipelines, and the unsexy plumbing in between.
Agents, retrieval, evaluation — or just comparing notes on what's actually working.
I lead AI integrations at a German enterprise consultancy, build multi-agent systems and audit pipelines, and quietly research low-resource language preservation on the side.
Outside of work I'm reading for an Executive Programme at Oxford and an MSc at Abertay. I think a lot about evaluation, governance, and the human bits we keep skipping.
Reliable multi-agent architectures: planners, tool use, eval harnesses, and the boring guardrails that keep them honest in production.
Hybrid search, re-rankers, chunking strategies — and the unsexy eval work that decides whether your bot can be trusted with a customer.
Taking notebook-shaped prototypes and turning them into deployable services with monitoring, drift detection, and clean rollback.
For teams sitting on a backlog of AI ideas: structured discovery, feasibility, scoping, and roadmaps that survive contact with reality.
Each one solved a stubborn problem — a low-resource language with no corpus, a research workflow drowning in PDFs, a legal text that nobody can quote correctly. Below, the longer version.
End-to-end framework preserving Igala, a Nigerian language with little digital corpus. Custom GloVe embeddings, bilingual translation API, dictionary, and synthetic data pipelines from bilingual sources.
Read case studyA research platform with a CLI-flavored interface: streaming markdown, scholarly source integration, document uploads, and interactive visualisations. Built for the kind of researcher who reads twelve papers a day.
Read case studyRAG system over Nigeria's full legal corpus with citation-grounded retrieval, source-traceable answers, and hallucination guards tuned for legal research. Built for lawyers who want receipts.
Read case studyPhotographed handwritten notes → cleaned, structured digital text via OCR + LLM error correction.
Cross-platform app with Gemini-powered summaries, voice input, and offline cache.
Python package combining Tesseract with GPT / Claude / Gemini for intelligent doc processing.
Comparative study: Naïve Bayes vs. CNN for detecting false articles. CNN wins decisively.
Image processing pipeline combining OCR with generative AI for visual data interpretation.
Interactive chatbot for UN Sustainable Development Goals — personalised, action-oriented.
Production sentiment classifier hitting 90%+ accuracy on social media corpus.
Twelve more public repos — language models, scrapers, evaluation harnesses, and side experiments.