AI Systems
The decisions behind the AI, not the models behind it
These examples connect model behavior to human judgment: interpreting a prediction, reviewing a draft, asking for better help, or inheriting a decision. Each links to its evidence and limitations.
Keeping people in control
When an AI can act, not just answer, what stops it acting on your behalf without your say-so?
Knowing when not to comply
What should a system do when giving someone exactly what they asked for works against them?
- flexi
When giving the user what they ask for is the wrong answer
The design question behind Flexi: what an AI should do when satisfying a request undermines the user's actual goal.
Open evidence - flexi
Visible choices after the first explanation
Follow-up actions offer simpler explanations, analogies, translation, detail, rephrasing, and challenges.
Open evidence
Being honest about limits
How does a system show the edge of what it knows, to someone with no way to check?
- ck12-analytics
Make uncertainty interpretable
The Foresights range view makes uncertainty visible instead of implying a precise single value.
Open evidence - ck12-analytics
Perceived value exceeded comprehension
Two published evaluations, ten teachers each using a demo class, found graph difficulty for half of participants.
Open evidence
Context that compounds
How does work accumulate instead of restarting from an empty box every time?
- offboard
Every opportunity is a persistent workspace
The central product decision: an opportunity is the unit the whole system is organised around, and context accumulates inside it.
Open evidence - offboard
Downstream artifacts inherit upstream work
Research feeds the resume, the resume and research feed outreach, and all of it feeds interview preparation — nothing restarts from empty.
Open evidence - offboard
LUMO as a context-aware layer, not a separate chatbot
The AI operates across the opportunity workspace rather than sitting beside it as an isolated chat surface.
Open evidence
Systems with more than one user
How do you hold competing needs together without quietly picking a winner?
- ck12-analytics
Separate prediction from diagnosis
Foresights supports planning before work; Insights supports investigation afterward.
Open evidence - experiment
Preserve the reasons behind an operator decision
Attach recommendations to assets, scale friction to reversibility, and retain defer or override reasoning.
Open evidence
Built, not just specified
What does the architecture actually look like underneath?