kit.ai

Inspiration

What it does

KIT AI is an offline‑first emergency first‑aid assistant designed for situations where internet access and medical professionals aren’t available. It provides objective, widely accepted first‑aid guidance using a locally running AI model and a vetted, versioned knowledge base. Once installed, KIT AI works entirely offline, helping users respond to emergencies on planes, during outdoor activities, or in low‑connectivity regions.

How we built it

We built KIT AI as a Progressive Web App (PWA) using React + Vite, with offline support via service workers and IndexedDB. A local LLM runs fully on‑device (WebGPU/WASM) to understand user questions and clearly rephrase pre‑approved first‑aid content.

Medical knowledge is stored as structured, versioned files that are cached locally. When internet access is available, a Gemini‑powered content pipeline generates and updates this knowledge base in the cloud. The app syncs these updates when online, then continues working offline without any cloud dependency during use.

Challenges we ran into

Offline constraints: Running an LLM entirely on‑device required careful model selection, caching, and performance tuning.

Safety boundaries: We had to strictly limit the AI so it could not diagnose, prescribe, or invent medical advice.

API quotas & reliability: Cloud AI quotas and access required us to design a system that never depends on live cloud calls during emergencies.

Scope control: Medical knowledge is massive, so we had to design a curriculum‑based approach to keep the system reliable and maintainable.

Accomplishments that we're proud of

A fully functional offline AI experience that works without internet.

A safety‑constrained architecture where AI cannot hallucinate medical advice.

A clean separation between online knowledge updates and offline usage.

A realistic, ethical approach to AI in emergency and health‑adjacent scenarios.

What we learned

Offline‑first design forces better engineering decisions.

AI systems in sensitive domains need clear boundaries, not just disclaimers.

Treating AI as a content compiler, not a free‑form chatbot, leads to safer and more reliable products.

Versioned data pipelines are just as important as model choice.

What's next for kit.ai

Expand the curriculum into modular “packs” (outdoor, travel, disaster response).

Improve offline retrieval and response speed.

Add multilingual support for global accessibility.

Continue refining safety, clarity, and UX for high‑stress emergency situations.