AidMate – Offline AI Health & Crisis Companion

Inspiration

During health or crisis situations, internet connectivity is often unreliable. We wanted to create a tool that works both online and offline, ensuring that people can still get vital health information and emergency alerts. AidMate was inspired by the need for a trustworthy AI companion that doesn’t stop working when the network does.

What it does

AidMate is an AI-powered health and crisis assistant that:

Collects and summarizes critical alerts from trusted sources.

Provides answers using retrieval-augmented generation (RAG).

Sends emergency SMS notifications via Twilio.

Works offline using a local LLM (Ollama) so users aren’t cut off during outages.

Displays alerts and summaries on a simple dashboard for quick access.

How we built it

Backend: FastAPI to orchestrate ingestion, RAG, and notifications.

AI/ML: FAISS vector store for embeddings, LangChain for orchestration, Ollama LLM for offline fallback.

Scheduler: GCP Cloud Scheduler for automated hourly data fetching.

Deployment: Containerized on GCP Cloud Run with secrets stored in GCP Secret Manager.

Notifications: Twilio integration for SMS alerts.

Dashboard/UI: Simple interface showing summarized alerts and contextual responses.

Challenges we ran into

Setting up a seamless offline fallback AI model without breaking user experience.

Managing vector storage and ensuring fast, accurate retrieval.

Handling deployment complexities across local dev and cloud production.

Avoiding hard-coded secrets and securing API keys with GCP Secret Manager.

Designing the system to be both lightweight and resilient.

Accomplishments that we're proud of

Built a hybrid online-offline AI assistant that works even without internet.

Created an automated alerts pipeline with summarization and notification features.

Designed a modular architecture that can easily extend to new data sources.

Deployed successfully on Google Cloud Platform with monitoring and logging.

What we learned

How to integrate offline AI models into production systems.

Best practices for secure deployments using GCP Secret Manager.

The importance of summarization for making critical information digestible.

Balancing cloud-native tools with local resiliency for crisis apps.

What's next for AidMate – Offline AI Health & Crisis Companion

Adding voice input and speech output for accessibility.

Expanding data sources for richer health and crisis alerts.

Training custom domain-specific health models for better accuracy.

Building a mobile-first experience for wider reach.

Adding provenance and explainability so users can trust every response.

Built With

custom
dashboard
faiss
faiss)
github-(version-control-&-ci/cd)
google-cloud-logging-&-monitoring-apis-&-integrations:-twilio-(sms-notifications)
google-cloud-scheduler
google-cloud-secret-manager
google-cloud-storage-(persistence)-cloud-services:-google-cloud-run
google-kubernetes-engine-(gke)
langchain
languages:-python-(fastapi
ollama-(offline-ai-model)-other-tools:-docker-(containerization)
ollama-llm-databases-/-storage:-faiss-(vector-store)
powershell-(setup-scripts)-frameworks-&-libraries:-fastapi
ui
web

Submitted to

OpenAI Open Model Hackathon

Created by

I led the backend design and implementation of AidMate. My contributions included:
• Building the FastAPI services that power the app, with endpoints for RAG-based question answering, text-to-speech, speech-to-text, and vector database ingest.
• Integrating OpenAI and Ollama models with dynamic logic to seamlessly switch between cloud (OpenAI) and offline (Ollama) depending on internet availability.
• Designing the data pipeline to fetch crisis alerts from sources like NWS, USGS, and AirNow, normalize them, and store them in FAISS vector DB for retrieval.
• Implementing session history and emergency flag detection, ensuring context-aware conversations and crisis keyword alerts.
• Engineering offline resilience so that the system continues to serve critical advice when connectivity is limited.

Through this work, I deepened my expertise in LangChain, FAISS, LLM orchestration, and distributed crisis data ingestion while focusing on building a reliable, real-time companion for emergency situations.

praveen srivastava
I led the infrastructure setup and deployment of a large-scale GPT OSS model on Google Cloud with a Windows-based environment. My role focused on selecting the right machine family, configuring resources, and troubleshooting VM startup failures caused by global CPU quota limits and GPU availability across regions.

A key challenge I overcame was navigating Google Cloud’s complex quota system — understanding the constraints of E2 vs A2 instances, unlocking GPU options in supported zones, and balancing resource needs against cost. I’m especially proud of resolving these limitations and creating a clear, repeatable path to run a 20B-parameter model on the cloud.

Through this project, I deepened my knowledge of cloud infrastructure scaling, GPU/CPU quota management, CUDA setup on Windows, and model deployment strategies for large AI systems.

Madhu Brahmanlda Pally
Shames tabrez khan
Thinker And Contributor
Shreyash Srivastva
i code
Cheetah11
Lalit Shandilya
Jitendra Kumar