Interdisciplinary researcher
An interdisciplinary researcher working across machine learning, security, and the applied sciences β from understanding what AI is actually doing to pointing it at problems that matter for people who are usually left out.
AI safety & interpretability β I want models we can actually inspect, not just score.
- Mechanistic interpretability with Stochastic Parameter Decomposition β mechanism clustering on real grokked transformers (
spd-real-transformer,spd-mechanism-clustering,spd-causal-clustering). - Cross-architecture universality of deception features (
deception-feature-universality). - Reward generalization early-warning (
reward-generalization-earlywarning). - Multilingual AI safety on Tiny Aya Earth for Yoruba, Hausa and Igbo (article,
lexsi-safety). Top 10% of 369 submissions at the Lexsi AI Safety and Alignment Hackathon 2026.
Agentic security β the reasoning process itself is becoming an attack surface.
reasoning-monitorβ detecting injected/absorbed premises in agentic reasoning chains, with activation probes.eval-oversightβ catching evaluation gaming with independent, held-out oversight.Teger-aiβ a browser & email plugin that highlights social-engineering cues, explains its reasoning, and builds an open dataset of phishing patterns. Supported by the OpenAI Cybersecurity Grant.
Applied AI for Africa β technology should reach the field, not just the paper.
- π±
Cropieβ crop-disease early warning Β·SoilIQβ soil-health intelligence Β·FarmVoice-AIβ agricultural extension agent. - π§ Water quality & availability monitoring (
water-quality-sentinel,Africa-water-watch). - πΊοΈ Blockchain-secured land registry & boundary-dispute resolution (
-LandChain-Africa).
And I ship product β full-stack apps, agent harnesses, and tooling around all of the above.
- Understand before you optimize. A score you can't explain isn't safety.
- Name the trade-off, then make the call.
- Build the thing, measure it honestly, and write down what still doesn't work.
- βοΈ stardukeza@gmail.com
- π¬ ORCID: 0009-0009-9311-2792
- π Google Scholar: Emmanuel Duke
- πΌ LinkedIn: emmanuel-duke
- π¦ X: @starduke001
- π» github.com/dukemawex
βTell me the goal, I'll handle the rest.β


