Skip to content

5-07/sweeten

Repository files navigation

sweeten

AI-Resilient Health Planning for Diabetes Management

Sweeten is a research-driven digital health system that investigates how hybrid AI and deterministic architectures can deliver safe, personalized diabetes support under real-world constraints — including API failure, data sparsity, and limited infrastructure.

Built on a failure-first design philosophy: the system must remain functional, interpretable, and clinically grounded even when AI is unavailable.

Live: sweeten-3aa2.vercel.app

Detailed Documentation: sweeten_narrative.pdf


Research Question

Can a health-planning system remain reliable and personalized even when AI components fail — and can the outputs of deterministic and AI-assisted pipelines be meaningfully compared?

Sweeten was built as a system-level experiment to answer this question.


Research Contributions

  • A hybrid AI + deterministic plan-generation pipeline with automatic graceful degradation
  • A shared analytics layer enabling direct, controlled comparison of AI vs. rule-based plan outputs
  • ADA guideline-to-code traceability — clinical thresholds mapped explicitly to conditional logic
  • Structured handling of real-world data sparsity — the system functions with partial weekly data
  • A reference architecture for AI-resilient health applications in resource-constrained settings

System Overview

User → Vitals Logging (calendar interface)
         ↓
   Weekly Aggregation + Analytics (planAnalytics.ts)
         ↓
   Plan Generation Pipeline
   ├── AI-Assisted: Gemini (gemini-2.5-flash), server-gated, rate-limited
   └── Deterministic Fallback: planEngine.ts, rule-based, always available
         ↓
   Weekly Health Plan (diet · activity · behavioral guidance)

Software Architecture

Frontend

Layer Technology
Framework Next.js 15 (App Router)
Language TypeScript
Styling Tailwind CSS
Auth Firebase Auth

The calendar-based vitals interface enforces structured, date-indexed data entry — mirroring real clinical logging behavior. UX design prioritizes low cognitive load and non-punitive framing.

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)

Data Model (Firestore)

Vitals stored per-user, per-day:

users/{uid}/vitals/{YYYY-MM-DD}
Field Type Required
glucose mg/dL
insulin units optional
carbohydrates grams optional
weight kg optional
steps count optional
mood enum optional
notes string optional

Schema intentionally supports partial and sparse data — a core aspect of the research problem.

Plan Generation Pipeline

The core research contribution: a two-tier architecture where the deterministic engine is primary and AI is an optional enhancement.

Tier 1 — AI-Assisted Generation

  • Implemented via server-side API routes (never client-side)
  • Model: gemini-2.5-flash
  • Enforced constraints: one plan per user per week, regeneration only on significant vitals change
  • AI refines tone, sequencing, and phrasing — it does not modify safety-critical logic
src/app/api/plan/    # AI plan generation endpoint
src/app/api/usage/   # Quota tracking and enforcement

Tier 2 — Deterministic Fallback Engine

Automatically activates on AI failure (quota exhaustion, error, or unavailability).

src/lib/planEngine.ts

The fallback engine:

  • Aggregates last 7 days of vitals
  • Computes trends and summary statistics
  • Applies ADA-grounded threshold classification
  • Generates structured weekly plans using conditional templates
  • Produces the same output schema as AI-generated plans

This ensures no hard system failure and continued personalization without AI.

Shared Analytics Layer

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

Both pipelines consume identical computed metrics:

  • Average glucose levels
  • Week-over-week glucose trend (improving / stable / worsening)
  • Activity level changes
  • Data completeness indicators

This shared foundation enables direct, controlled comparison between AI-generated and deterministic plans — a key evaluation capability.


Guideline-to-Logic Traceability

Every threshold in the plan engine maps to a published clinical guideline:

ADA Guideline Threshold / Trigger Plan Focus
Postprandial glucose target avg BG > 180 mg/dL Glucose Stability Focus Week
Physical activity ≥150 min/wk < 5,000 steps/day Activity Rebuild Week
Medication consistency insulin consistency score Adherence & Routine Week
Glycemic variability management high SD in BG readings Dietary Carb Awareness Week

Evaluation Approach

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

  • Robustness testing — AI availability simulated as present, degraded, and absent
  • Behavioral consistency — AI and deterministic plan outputs compared on structure and focus selection via shared analytics
  • Failure simulation — missing data, extreme glucose values, zero-entry weeks
  • Qualitative analysis — plan completeness, tone, and clinical alignment against ADA mappings

Key Findings

Deterministic and AI-generated plans agreed on plan focus selection in the majority of tested cases, but diverged on tone, sequencing, and recommendation specificity. This suggests the AI layer's primary value is communication quality, not correctness — with implications for how hybrid health systems should be evaluated.


Safety and Ethics

  • Not a medical device; no diagnostic or clinical use
  • No medication dosing logic
  • AI never invoked client-side; all calls server-gated
  • Weekly limits prevent over-generation
  • Data privacy enforced via Firestore security rules
  • No personal health data used for model training

Why This Is Research

Sweeten explicitly studies:

  • AI failure modes in health-adjacent software systems
  • Hybrid intelligence architectures and graceful degradation strategies
  • Ethical constraints on AI usage in healthcare contexts
  • Clinical guideline operationalization in rule-based systems

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


Future Directions

  • Clinician-validated evaluation of plan clinical alignment
  • Localized guideline adaptation for South Asian metabolic profiles
  • Formal ablation study: AI vs. deterministic plan quality metrics
  • Integration with glucometers and consumer wearables
  • Extension of this architecture to other chronic condition domains

Stack

Next.js 15 · TypeScript · Tailwind CSS · Firebase Auth · Firestore · Google Gemini


Status

Solo-developed research system. Built for science fairs, research showcases, and academic expansion.

Developer & Researcher: Muzaina Munir


Sweeten is a decision-support and habit-structuring tool. It is not a medical device and does not provide diagnostic or clinical advice.

Releases

No releases published

Packages

 
 
 

Contributors

Languages