A custom AI built to honour a mathematical mind living with early dementia

Project Description TJ’s Theoretical Studies: Identity-Aligned AI Interaction Architecture Achievements I designed and engineered a behaviourally constrained AI interaction architecture for a cognitively vulnerable yet intellectually capable individual experiencing early-stage dementia. The intended user has a lifelong identity rooted in advanced mathematics and structural reasoning. While dementia has begun affecting planning capacity and introducing paranoid ideation, high-level analytical reasoning remains intact and central to self-concept. Rather than deploying simplified puzzles, gamified activities, or conversational distraction, I architected a Custom GPT that generates rigorous mathematical “Studies” requiring pen-and-paper reasoning. The system intentionally withholds full derivations unless explicitly confirmed, preserving agency and reinforcing intellectual ownership. Layered behavioural guardrails prevent exposure to high-risk thematic domains and safely redirect suspicious or paranoia-related inputs without validation or escalation. This project is not a prompt experiment. It is a structured interaction architecture engineered around a specific human identity profile. Tools - ChatGPT Custom GPT Builder - Structured instruction design - Behavioural systems modelling - Iterative prompt architecture refinement - Multi-round simulation and boundary stress testing The entire system was engineered within GPT’s instruction framework using layered behavioural constraints rather than external automation or plug-ins. Methodology The design process followed five structured phases: 1. Identity and Constraint Definition Mapped preserved cognitive strengths (advanced analytical reasoning) and vulnerabilities (planning deficits, fixation risk, paranoid ideation). Defined system objectives: preserve dignity, maintain agency, and align interaction with identity. 2. Risk Mapping and Domain Control Identified high-risk narrative domains that could reinforce distress or fixation, including: - Financial themes - Surveillance framing - Health scenarios - Personal contextual references - Threat or scarcity framing These categories were explicitly excluded through instruction-level guardrails. 3. Layered Interaction Architecture Design Engineered the GPT as a behavioural system with containment layers: - Domain rotation logic to maintain novelty - Solution-withholding and escalation-gate rules to preserve cognitive ownership - Anti-triviality safeguards to prevent anticlimactic or simplistic problems - Controlled acknowledgement language to avoid patronising tone - Suspicion protocol for neutral redirection of monitoring-related prompts - Accessibility calibration using plain-text mathematical notation The system was designed not to optimise for answer delivery, but for structured reasoning engagement. 4. Failure Mode Stress Testing Conducted adversarial simulations to test boundary behaviour under edge conditions, including: - Equation leakage - Over-assistance - Tone drift - Verbosity inflation - Guardrail breaches - Suspicion or monitoring prompts - Silence and hesitation misclassification Each behavioural layer was refined until containment and response discipline remained stable under repeated stress. 5. Pre-Deployment Calibration Calibrated difficulty progression, acknowledgement tone, escalation rules, and closing logic prior to live rollout. The GPT was treated as a dynamic behavioural architecture requiring containment, calibration, and iterative refinement — not as a static content generator. Results (Pre-Rollout Validation) Simulation testing confirms: - Guardrails activate reliably under suspicious prompts. - The model withholds full derivations unless explicitly confirmed. - Tone remains neutral, structured, and non-patronising. - Interaction avoids escalation pathways. - Topic drift and silence are handled through defined protocols. - Agency is preserved by preventing premature solution completion. The rollout is guided by a defined hypothesis: Identity-aligned intellectual engagement may provide structured diversion and preserve agency during cognitive fluctuation. Deployment will proceed in controlled phases with iterative refinement. Learnings This project reframed AI from prompt engineering to interaction architecture. Key insights: - Safety must be engineered in layers; it cannot be assumed. - AI systems tend to overperform by completing tasks prematurely, which can remove agency. Intentional constraint preserves dignity. - Behavioural guardrails must be stress-tested against adversarial inputs, not simply described. - Tone discipline is as important as functional correctness. - Meaningful AI impact emerges when systems are designed around a specific human identity profile rather than a generic user category. The broader implication is that highly individualised AI interaction architectures may offer safer and more meaningful engagement for vulnerable populations when designed around preserved strengths rather than deficits. Future Vision This project represents an early prototype of identity-aligned AI interaction systems. Rather than universal assistants, future AI architectures may function as configurable cognitive environments that evolve with an individual’s psychological and cognitive profile. Such systems could support populations experiencing: - Cognitive decline - Neurodivergence - Anxiety disorders - Identity disruption Beyond direct users, personalised AI architectures may reduce strain on families and carers by providing structured, safe engagement that adapts over time. The opportunity is not merely productivity. It is dignity-preserving, identity-aligned human-AI interaction. Relevant URLs Custom GPT deployment is private to protect user safety and privacy. Access can be provided upon request for evaluation purposes.