Turning complex public-service communications into
structured case intelligence.

THE PROBLEM

Important case information is often scattered across institutions and messages.

Residents dealing with councils, health services and other public bodies can face fragmented communication, difficult-to-track events and uncertainty about the published duties or standards relevant to their situation. CareConnect explores how AI can organise that information without replacing the user's judgement.

CONTEXTUAL ONBOARDING

Build the case context before asking the model to reason.

• Borough-specific onboarding for Hackney and Lambeth.
• Selection of relevant institutions such as council housing, social services, NHS services and Universal Credit.
• Optional personal-context signals used to identify potentially relevant published rights and duties.
• Clear separation between case setup, rights discovery, communications import and timeline review.

AI RIGHTS REPOSITORY

Structured regulatory information instead of unbounded chat.

The Gemini service uses a constrained UK public-law system instruction covering legislation and local-service standards. The application requests machine-readable JSON containing right titles, summaries, sources and tags, then renders those results into a dedicated rights repository rather than exposing raw model output.

TIMELINE EXTRACTION

Convert communications into reviewable factual events.

• Normalised message objects for Microsoft or Google-style communication sources.
• Gemini extraction of timeline events and suggested next actions.
• Typed categories including decisions, meetings, information and actions.
• Confidence levels attached to extracted events.
• Draft events remain unconfirmed until the user explicitly accepts them.
• Confirmed events are separated from AI-generated drafts in the case timeline.

ENGINEERING DETAIL

Schema-constrained AI inside a typed frontend workflow.

CareConnect uses the Google GenAI SDK with explicit response schemas rather than relying on free-form text parsing. TypeScript models define institutions, vulnerabilities, rights, source messages, timeline events, actions and onboarding state. This creates a predictable boundary between probabilistic model output and deterministic application state.

PROTOTYPE BOUNDARY

Designed honestly around what is implemented.

The current repository demonstrates the product flow and AI reasoning layer. The Outlook/Gmail connection surface uses a simulated message import in this version, so production mailbox OAuth, secure server-side ingestion, persistent storage, consent logging and privacy controls would be the next engineering layer before handling live user communications.

WHAT THIS DEMONSTRATES

AI product design with domain constraints and human review.

The project demonstrates contextual AI application design, schema-constrained LLM outputs, typed state modelling, domain-specific prompting and human-in-the-loop review for information that needs more care than a generic chatbot.