Sweeten

AI-Resilient Health Planning for Diabetes Management

Sweeten is a research-driven health planning system that investigates how hybrid AI and deterministic architectures can maintain safe, personalized diabetes support under real-world constraints such as API limits, missing data, and system failure.

Rather than assuming constant AI availability, Sweeten is built around a failure-first design philosophy: the system must remain functional, interpretable, and useful even when AI services are unavailable.

---

Table of Contents

• Research Motivation
• Research Contributions
• System Overview
• Software Architecture
  • Frontend
  • Data Model (Firestore)
  • Plan Generation Pipeline
    • AI-Assisted Generation
    • Deterministic Fallback Engine
  • Analytics Layer
  • Persistence and Safety Controls
• Evaluation Approach
• Why This Is Research
• One-Sentence Pitch
• Status

Research Motivation

Many modern health applications rely entirely on large language models for personalization. In low-resource or high-constraint environments, this introduces critical failure modes:

• API quota exhaustion or downtime
• Inconsistent outputs under sparse or incomplete data
• Unsafe degradation when AI is unavailable

Research question:

Can a health-planning system remain reliable and personalized even when AI components fail?

Sweeten was designed as a system-level experiment to explore this question.

---

Research Contributions

• A hybrid AI + deterministic plan-generation pipeline
• Graceful degradation strategies for AI failure in healthcare contexts
• Structured handling of real-world data sparsity
• A reference architecture for AI-resilient health applications

Sweeten is explicitly framed as a decision-support system, not a diagnostic or clinical tool.

---

System Overview

User → Vitals Logging → Weekly Aggregation ↓ Plan Generation Pipeline (AI-assisted → deterministic fallback) ↓ Weekly Health Plan

---

Software Architecture

Frontend

• Framework: Next.js 15 (App Router)
• Language: TypeScript
• Styling: Tailwind CSS
• Authentication: Firebase Auth (client-side)

The frontend is responsible for vitals entry, plan visualization, and enforcing UX-level constraints such as weekly plan limits.

src/app/ ├── dashboard/ # Plan preview and generation entry point ├── vitals/ # Calendar-based vitals logging interface ├── plan/ # Displays latest generated weekly plan └── api/ # Server-side routes (AI, usage, persistence)

The calendar-based vitals interface enforces structured, date-indexed data entry, mirroring real clinical logging behavior.

---

Data Model (Firestore)

Vitals are stored per-user, per-day using a deterministic document structure: users/{uid}/vitals/{YYYY-MM-DD}

• Required: glucose (mg/dL)
• Optional: insulin, carbohydrates, weight, blood pressure, steps, mood, notes

This schema intentionally supports partial and sparse data, a key aspect of the research problem.

---

Plan Generation Pipeline

The core research contribution of Sweeten is its two-tier plan generation architecture.

AI-Assisted Generation

• Implemented via server-side API routes
• Uses Google Gemini (gemini-2.5-flash)
• Enforced constraints:
  • One plan per user per week
  • Regeneration only on significant vitals change
• AI is never invoked client-side

src/app/api/plan/ # AI plan generation endpoint src/app/api/usage/ # AI usage and quota tracking

This design enables controlled experimentation with AI availability, cost, and reliability.

---

Deterministic Fallback Engine

When AI services fail (quota exhaustion, error, or unavailability), Sweeten automatically falls back to a rule-based planning engine.

src/lib/planEngine.ts

The fallback engine:

• Aggregates weekly vitals
• Computes trends and summary statistics
• Generates structured weekly plans using deterministic logic
• Produces the same output schema as AI-generated plans

This ensures:

• No hard system failure
• Predictable, explainable behavior
• Continued personalization without AI dependency

---

Analytics Layer

Shared analytics logic is used by both AI and fallback pipelines.

src/lib/planAnalytics.ts src/lib/planEngine.ts

Computed metrics include:

• Average glucose levels
• Week-over-week glucose trends
• Activity changes
• Data completeness indicators

This allows direct comparison between AI-generated and deterministic plans.

---

Persistence and Safety Controls

• Generated plans are stored in Firestore
• Weekly limits prevent over-generation
• No real-time recommendations
• No medication dosage adjustments
• No emergency or diagnostic logic

Sweeten prioritizes system safety and predictability over real-time intervention.

---

Evaluation Approach

Sweeten is evaluated as a software system, not a medical treatment.

• Robustness testing under AI availability and failure
• Behavioral consistency between AI and fallback outputs
• Failure simulations with missing or extreme data
• Qualitative analysis of plan completeness and tone

This enables meaningful evaluation without clinical trials.

---

Why This Is Research

Sweeten explicitly studies:

• AI failure modes in health systems
• Hybrid intelligence architectures
• Graceful degradation strategies
• Ethical constraints on AI usage in healthcare

The codebase is structured to support experimentation and analysis, not just deployment.

---

One-Sentence Pitch

Sweeten is a research-driven health planning system that explores how hybrid AI and deterministic architectures can deliver safe, reliable diabetes support under real-world constraints.

---

Status

Sweeten is a solo-developed project intended for:

• science fairs
• research showcases
• hackathons
• further academic expansion