AegisCARE: OFFLINE AI FIRST RESPONDER

Inspiration

AegisCARE was inspired by a simple fear: emergencies don’t wait for internet. I have seen network bars vanish during floods, travel, or blackouts while the situation keeps getting worse. I asked, “Why does lifesaving help still depend on the cloud?” That question drove me to design something that works when towers fall, Wi-Fi fails, and privacy matters most.

What it does

AegisCARE turns a smartphone into an offline AI first responder. It listens to speech locally, reasons using small on-device language models, and acts through agentic tools only after user approval. It can guide CPR, bleeding control, stabilization, call emergency numbers, notify relatives, draft short medical summaries, and store encrypted records, all without internet.

How I built it

I started by designing the system around one rule: everything must run on-device. The pipeline begins with the microphone, where speech is captured locally and converted to text using a fully on-device Whisper speech-to-text transformer. No audio is uploaded at any stage. The refined text is then routed through the RunAnywhere SDK, which loads and executes quantized Small Language Models (DeepSeek-R1-Distill-1.5B or Llama-3-3B in Q4/Q5). RunAnywhere manages model loading, hardware acceleration, and routing between STT → SLM → TTS while keeping memory use practical on phones. I then designed an agentic workflow. The SLM does not just answer; it plans actions. Based on local tools, it can suggest calling emergency services, notifying relatives, fetching nearby hospitals, prioritizing first-aid actions, or generating a short summary. Tools execute locally, and actions only proceed after user approval. The response is spoken back using local text-to-speech, and notes are encrypted and stored on the device. Quantization and on-device inference optimization were essential to manage heat, battery, and latency. In short, AegisCARE is a self-contained agent on the phone, not a thin client for someone else’s server.

Challenges I ran into

A major challenge was understanding what it really means to run DeepSeek-R1-Distill and Llama-3 fully on-device rather than in the cloud. I had to study quantization options, memory limits, and whether a 1.5B to 3B parameter model can realistically run on a 6 to 8 GB RAM phone without overheating. Designing an agentic system around DeepSeek was also a conceptual challenge. The model is powerful, but deciding which actions it should be allowed to execute offline, such as calling numbers, reading medical notes, or guiding procedures, required careful thought about safety and user control. Since I was not writing code, the problems I faced were mostly intellectual ones, such as mapping real emergencies to agent workflows, deciding when DeepSeek should reason versus defer, and imagining failure cases when there is no server and no internet to rescue the system.

Accomplishments that I'm proud of

I am proud that AegisCARE has a complete, realistic architecture rather than just an idea sentence. I created detailed data-flow diagrams, selected feasible on-device models, and designed an agentic workflow that could actually run on today’s phones. The project involved reading research papers, model specifications, and edge-AI deployment constraints instead of simply wiring APIs together. The accomplishment here is clarity: I now have a design in which every component : Whisper, RunAnywhere SDK, quantized SLMs, on-device routing, and encryption has a defined role and justification.

What I learned

I learned how different on-device AI is from cloud AI, both technically and philosophically. I studied how quantization allows large models to fit on mobile hardware, how agentic systems plan actions instead of only answering questions, and how offline-first design completely changes the user experience. More importantly, I learned to evaluate feasibility instead of assuming that the cloud will handle everything. Even though the project is theoretical, the learning was very real, and I realized that architecture matters just as much as code.

What's next for AegisCARE: OFFLINE AI FIRST RESPONDER

I already have a working prototype, so the next step is to move toward deployment. My focus now is testing the prototype on different phones, measuring latency, battery impact, and overheating behavior, and refining model size and quantization choices so it remains reliable during long offline use. I also want to conduct small user trials to see how people actually interact with an agentic offline assistant during stressful situations.

Beyond deployment, the roadmap is broader than healthcare. The same architecture can power offline safety, accessibility, education, agriculture, construction, family assistance, and drone or field analytics, all without sending any data to the cloud. I plan to gradually add modules such as camera-based scene narration, PPE safety checks, multilingual support, pediatric guidance, offline family AI hub features, and rural assistance tools.

Another next step is building strong safety guardrails around what the agent is allowed to do autonomously while keeping the user in full control. Longer term, I want AegisCARE to evolve from a single app into a general on-device agent platform.

Built With

• agent
• ai
• javascript
• python
• runanywheresdk