KiokuNet-A safety Companion

Inspiration

Small changes in an older adult’s day—missing a routine, feeling lost, or becoming unusually quiet—can be difficult for families to interpret. I wanted to build something that offers calm support to the older adult while giving caregivers meaningful context instead of sending automatic, unexplained alerts.

What it does

KiokuNet is an AI safety companion for older adults and their caregivers. It has a patient view for Meera and a caretaker view for Ananya.

Meera can ask Nia about medication reminders, mark medicine as taken, speak or type messages, share a photo, and receive calm support. The caretaker can review evidence-linked safety plans, medication status, safe-zone context, explainable drift factors, and a daily care journal.

KiokuNet uses GPT-5.6 to reason over Meera’s persona, baseline routine, recent behavior, surfaced deviations, direct messages, and optional image context. It returns a structured decision, confidence, cited evidence, immediate action, follow-up action, and caregiver summary.

How I built it

I built KiokuNet with React, Vite, Node.js, Express, Leaflet, and the OpenAI Responses API. The project uses the GPT-5.6 model configured as gpt-5.6-luna.

The backend first creates deterministic behavior data and surfaces factual expected-versus-actual deviations. GPT-5.6 is then used as the reasoning layer, rather than as a simple classifier. It can use function calls to check medication schedules, create or acknowledge reminders, remember a stated destination, and escalate to a caregiver when the evidence supports it.

I used Codex throughout the build to turn a phased product plan into the working application, implement the shared patient/caretaker state, improve the interface, build the GPT tool-calling flows, and add tests and evaluation policies.

Challenges I ran into

The main challenge was making the AI helpful without making it overconfident. A routine deviation alone should not automatically mean an emergency.

To address this, I separated factual deviation detection from GPT reasoning. The model must cite surfaced evidence in its reasoning, follow a strict JSON schema, and avoid inventing facts. I also added validation, a retry for malformed outputs, and a safe fallback response.

Another challenge was making the patient experience feel reassuring. Nia avoids technical language and does not repeatedly ask Meera to share her location, because location sharing is already part of the care flow.

Accomplishments that I’m proud of

I am proud that KiokuNet is more than a dashboard or chatbot. It connects a patient companion, caregiver dashboard, behavior timeline, medication reminders, safe-zone context, photo verification, and AI-generated care journal in one shared experience.

I am also proud of the evaluation harness. It accepts more than one reasonable model decision when evidence supports it, but rejects unsupported reasoning. For example, continued monitoring after prolonged stillness is accepted only when GPT-5.6 cites later recovery evidence.

What I learned

I learned that agentic AI systems need strong boundaries around what the model sees, what it can do, and how its output is validated. GPT-5.6 is most useful here when it reasons over structured context and can use tools for real actions, rather than producing a generic conversational response.

I also learned how Codex can accelerate an end-to-end build: planning, implementation, debugging, testing, documentation, and iteration all happened much faster with an agentic development workflow.

What's next for KiokuNet

Next, I would add secure authentication, persistent encrypted storage, caregiver notification channels, configurable consent controls, multilingual support, and more real-world wearable or health-device integrations.

I would also expand the evaluation dataset with more diverse routines and edge cases, while keeping the core principle unchanged: KiokuNet should support human caregivers with context, compassion, and accountability—not replace them.