Social workers are the backbone of our communities, but they are trapped in a digital dark age. Up to 65% of their week is spent fighting fragmented paperwork, unsearchable PDFs, and chaotic legacy systems instead of actually helping people in crisis. When cases change hands, critical history is lost.

Waypoint is a mobile-first, AI-powered case memory platform designed to solve this. It doesn't just digitize notes—it utilizes advanced LLM architectures to act as a secure, continuous intelligence layer for case workers on the go.

Waypoint is built to be used in the field, primarily accessed via mobile devices while workers move between client visits.

• Crisis-Aware Dashboard: Workers instantly see their assigned clients, automatically tagged with risk levels (e.g., HIGH RISK: EVICTION) derived from their recent case history.
• Instant Voice Ingestion: Workers can record a disorganized voice memo immediately following a client visit. The platform transcribes the audio and restructures it into an objective, subpoena-safe, factual case note.
• Thread-Scoped Case Memory (RAG): Workers can upload legacy case documents (PDFs, court notices) directly to a client's profile. They can then chat with the client's localized history, instantly retrieving answers based on years of collected data.
• Audio Recaps (ElevenLabs): Because workers drive between visits, they can tap a button to hear a highly concise, AI-generated voice recap of the client's current crisis. They get crucial context safely, before ever stepping out of the car.

We built Waypoint around two core technologies, treating AI not as a gimmick, but as an orchestration layer and strict API boundary.

We didn't just use Backboard as a simple LLM wrapper; we leveraged its deep memory features and custom model routing:

• Dynamic Dual-Model Routing: We hooked into Backboard's native Google integration. Lightning-fast ingestion of messy notes is routed to Gemini Flash (acting as an ETL pipeline). The resulting case thread is then passed to Gemini Pro for deep reasoning and risk assessment.
• Thread-Scoped RAG: Every client is mapped 1:1 with a unique Backboard Thread ID. When a document is uploaded, we utilize Backboard's native Thread-Level Document API. This ensures the RAG pipeline is strictly scoped to that individual client, totally preventing cross-client data leaks.

Security is non-negotiable for HIPAA and PIPEDA-compliant data.

• Universal Login: We use Auth0 for seamless passwordless/MFA authentication.
• Strict API Gating: Auth0 sits entirely at our backend API boundary. Every time our frontend requests an AI action (like ingesting a transcript or summarizing a file), an Auth0 access token validates the request. We utilize the 'Secure AI Agents' pattern, ensuring AI operations only execute on behalf of a verified user, paving the way for human-in-the-loop push approvals for high-risk data mutations.

• Frontend: Next.js (App Router), React, Tailwind CSS, shadcn/ui
• Authentication & Security: Auth0 (@auth0/nextjs-auth0)
• AI & Orchestration: Backboard.io (Agentic Memory), Google Gemini (Flash & Pro via native integration)
• Database & Storage: Supabase (PostgreSQL, Storage Buckets)
• Voice & Audio: Native Web Speech API (Transcription), ElevenLabs (Voice Generation)

First, clone the repository and install dependencies:

npm install

Ensure you have your .env.local configured with the required keys for Supabase, Auth0, Backboard, and ElevenLabs.

Run the development server:

npm run dev
# or
yarn dev
# or
pnpm dev
# or
bun dev

Open http://localhost:3000 with your browser to see the application in action.