Skip to content

PROJECT_PURPOSE.md

Latest commit

 

History

History
228 lines (180 loc) · 8.96 KB

PROJECT_PURPOSE.md

File metadata and controls

228 lines (180 loc) · 8.96 KB

Bandicoot: Open-Source RMAB Platform for Health Outreach

Why Are You Building This?

You're building Bandicoot to solve a critical problem in global health: how to maximize impact when resources are limited.

The Problem

Health NGOs working on maternal and child care (like Suvita in India) face a fundamental challenge:

  • They serve hundreds of thousands of caregivers across multiple states
  • They have limited health workers for personalized outreach
  • 30-40% of beneficiaries drop out of engagement programs
  • SMS reminders alone aren't enough - but they can't call everyone
  • Every contact must count - but current methods use random or simple heuristics

The Result: Lives are lost because health workers can't prioritize who needs help most urgently.

The Proven Solution: Restless Multi-Armed Bandits (RMAB)

Google Research and ARMMAN NGO proved this works in India:

  • 30% reduction in dropout rates (from ~35% to ~24%)
  • Deployed at scale: 100,000+ beneficiaries
  • Now in production as SAHELI - the first RMAB system in public health
  • Scaled to 3 million mothers nationally

How it works: The system learns which mothers/caregivers are most likely to benefit from a call or SMS, then prioritizes limited health worker time to maximize engagement and health outcomes.

Your Vision: Make This Available to Every NGO

The problem: While Google/ARMMAN proved this works, it's not accessible to most NGOs.

Bandicoot will be:

  1. Open-source - Free for any NGO to use
  2. Self-hostable - Deploy on their own infrastructure
  3. Delivery-agnostic - Works with any SMS/call provider (Twilio, Pub/Sub, Kafka)
  4. Designed for scale - Handle millions of beneficiaries cost-effectively
  5. Built from real-world learnings - Leveraging SAHELI's successful deployment

The First Use Case: Suvita

Suvita is your reference implementation:

  • Mission: Improve childhood vaccination rates in India
  • Scale: 200,000+ caregivers across 7 states (Bihar, Maharashtra, UP, Rajasthan, West Bengal)
  • Approach:
    • SMS reminders 1 and 7 days before vaccine appointments
    • 5,000 → 35,000 volunteer "immunization ambassadors"
  • Challenge: Limited resources, declining engagement, ~25-30% of infants not fully immunized

Current Impact (without RMAB):

  • SMS alone: +2pp vaccination rate ($0.40-$1.71 per child)
  • Ambassadors: +7pp ($1-$2 per child)

Potential with Bandicoot:

  • 25-35% reduction in dropouts (based on ARMMAN results)
  • 20-60% improvement in health worker efficiency
  • Smarter allocation of SMS and ambassador resources

Technical Approach

Core Innovation: Contextual + Restless Hybrid

Combine two powerful techniques:

  1. Restless Bandits (bayesianbandits) - Model how engagement states evolve even without intervention
  2. Contextual Features (MABWiser) - Use demographics, location, history to personalize

Key Design Decisions (Inspired by SAHELI)

  1. Clustering - Group similar caregivers (~20 clusters) to share learning
  2. Whittle Index - Compute priority scores for each caregiver
  3. Features-Only Mapping - Assign new caregivers to clusters immediately
  4. Warmup Period - 6 weeks of data before RMAB prioritization kicks in
  5. Cost-Optimized - Serverless (Cloud Run), reuse existing infra, ~$100-200/month for 200K caregivers

Architecture

┌─────────────────┐
│ Caregiver Data  │
│ (SMS, Calls,    │
│  Vaccinations)  │
└────────┬────────┘
         │
         ▼
┌─────────────────────────────────────┐
│     Bandicoot RMAB Service          │
│  (FastAPI + bayesianbandits)        │
├─────────────────────────────────────┤
│ /train_clusters                     │
│ /precompute_whittle_indices         │
│ /assign_caregiver_cluster           │
│ /update_caregiver_state             │
│ /recommend?budget=K                 │
└─────────────────────────────────────┘
         │
         ▼
┌─────────────────┐
│ Top-K Priority  │
│   Caregivers    │
└────────┬────────┘
         │
    ┌────┴────┐
    ▼         ▼
  SMS    Ambassador
         Calls

Implementation Phases

Phase 1: Proof-of-Concept (Weeks 1-2)

  • Colab/Jupyter notebook with synthetic data
  • Validate RMAB mechanics
  • Demonstrate 20-30% improvement vs random allocation
  • Deliverable: Working demo showing engagement lift

Phase 2: MVP with Suvita (Weeks 3-6)

  • FastAPI service with /infer and /recommend endpoints
  • Clustering-based parameter learning (SAHELI approach)
  • Integration with Suvita's SMS and ambassador system
  • A/B testing vs current methods
  • Deliverable: Production-ready service showing real impact

Phase 3: Contextualization & Scale (Months 2-3)

  • Add demographic and temporal features
  • Hybrid bayesianbandits + MABWiser
  • Off-policy evaluation (IPS, DR estimators)
  • Deliverable: Enhanced model with personalization

Phase 4: Open Source Release (Month 4+)

  • Complete documentation
  • Terraform blueprints for GCP/AWS/Azure
  • Helm charts for Kubernetes
  • Example notebooks and datasets
  • Community governance model
  • Deliverable: Replicable platform for any NGO

Expected Impact

For Suvita:

  • 25-35% reduction in vaccination dropouts
  • 20-60% more efficient health worker utilization
  • Serve same 200K caregivers at ~$100-200/month infrastructure cost

For the Ecosystem:

  • Enable any health NGO to deploy proven AI optimization
  • Save thousands of lives by improving vaccination rates globally
  • Create open-source standard for RMAB in social impact
  • Build community of practice around AI for health

Technical Stack

Core Libraries:

  • bayesianbandits - Restless bandit support, delayed rewards
  • MABWiser - Contextual features, hyperparameter tuning
  • scikit-learn - Clustering, feature engineering
  • SAHELI's Whittle solver (open-source reference)

Infrastructure:

  • FastAPI - REST API
  • PostgreSQL - Persistent state
  • Redis - Fast lookups
  • Cloud Run - Serverless deployment
  • Pub/Sub/Kafka - Event-driven architecture
  • Docker + Kubernetes - Containerization
  • Terraform - Infrastructure as Code

Cost Optimization:

  • Serverless-first (pay only when used)
  • Preemptible VMs for batch jobs
  • Reuse existing databases
  • Free tier quotas (Cloud Run, Firestore)
  • Estimated: $100-200/month for 200K users

Why "Bandicoot"?

(To be determined - but the alternative name "Parvai" was considered, meaning "vision" or "sight" in Tamil, with the clever AI reference at the end)

Research Foundation

Built on peer-reviewed research:

  1. SAHELI/ARMMAN Field Study (AAAI 2022) - 30% engagement improvement
  2. Decision-Focused Learning (Wang et al.) - Train directly for policy optimization
  3. BCoR (Liang et al.) - Bayesian contextual RMABs for non-stationary behavior
  4. TARI (Danassis et al.) - Non-Markovian policies, 90% improvement
  5. WHIRL (Jain et al. 2024) - Inverse RL for reward learning
  6. GROUPS (Killian et al.) - Robust group-level planning

Success Metrics

Technical:

  • Cumulative regret vs optimal policy
  • Top-K hit rate for high-value interventions
  • Convergence speed of learning algorithm
  • Policy stability over time

Operational:

  • Vaccination attendance rate improvement
  • Health worker calls-per-success ratio
  • Cost per additional vaccination
  • Dropout rate reduction

Impact:

  • Lives saved (children fully vaccinated)
  • Cost-effectiveness ($/child vaccinated)
  • Scalability (beneficiaries served)
  • Replicability (NGOs adopting Bandicoot)

Beyond Suvita

Once proven, extend to:

  • Maternal health (antenatal care, postnatal visits)
  • TB treatment adherence
  • Nutrition counseling
  • Any health program with limited outreach resources

Your Bigger Picture

This appears to be part of a larger concern about preserving human agency in the face of economic and geopolitical pressures. By building open-source technical infrastructure that empowers communities and NGOs (rather than extractive tech monopolies), you're creating tools that enhance rather than erode human participation in shaping health outcomes.

In Summary

You're building Bandicoot to:

  1. ✅ Save lives by optimizing health outreach with proven AI
  2. ✅ Democratize access to cutting-edge RMAB technology
  3. ✅ Empower NGOs with self-hostable, open-source tools
  4. ✅ Prove that AI can augment (not replace) human health workers
  5. ✅ Create a replicable blueprint for social impact at scale

The goal: Transform how resource-constrained health programs operate globally, starting with Suvita's 200,000 caregivers and scaling to millions.