Offline-First Multi-Agent Resilience for the Mbagathi Basin
Project Lead: Mitchell Odili
Status: ✅ Level 3 Complete | 🌊 Level 4 Coordination Starting
Mission: Validating "Digital Guardian" protocols for urban flood resilience using multimodal LLMs and geospatial agent orchestration.
During the March 2026 rains in Nairobi, flash floods turned arterial roads (like Lang'ata and Mbagathi) into death traps within minutes.
- The Gap: Existing navigation tools (Google Maps/Waze) rely on active internet and crowdsourced data. When the grid fails, the data stops.
- The Impact: Residents lack the "Local Truth" needed to find safe ridges and avoid submerged underpasses, leading to preventable loss of life.
- The Reality: In a crisis, the map is often the first thing to go dark.
FloodPulse is an offline-first, multimodal resilience assistant. It uses the phone's native hardware to "see" and "reason" about flood risks without needing a cloud connection.
- Vision: A "Digital Guardian" for the Global South that works when the world goes dark.
- Validation: A Modular Agentic Simulation modeling the Mbagathi River basin to validate multi-agent consensus in high-risk scenarios.
We conducted a zero-shot analysis using Gemma 4 (31B) on high-resolution satellite imagery to validate core spatial reasoning.
- River Path: Successfully identified the riparian corridor despite urban canopy cover.
- Critical Nodes: Pinpointed three high-risk intersections:
- Lang'ata Road/ICC Crossing: Identified as a primary arterial bottleneck.
- South B/C Border: Identified as a "low-water" neighborhood split-point.
- Lower Basin Sumps: Corrected identified as high-risk vehicle entrapment zones.
- Safe Ridge Logic: The model autonomously identified the South B Plateau as a primary evacuation zone based on spectral terrain analysis (elevation vs. drainage).
Status: ✅ Feasibility Confirmed. The model demonstrates the required spatial intuition for urban flood navigation.
To simulate Nairobi's flood dynamics, we've architected a Parametric Persona Engine. This ensures that our agents—the Stranded Commuter, the Local/Boda Responder, and the Urban Strategist—have consistent identities.
Technical Implementation:
-
Orchestrator (
create_identity.py): The "Brain." Manages batch generation, directory routing, and Credit Saver logic for cost-efficient, idempotent runs. -
Worker (
generator.py): The "Muscle." Leverages Gemini 2.5 Flash chat sessions to maintain Visual Identity Consistency (e.g., ensuring a specific Kenyan flag beaded bracelet carries from portraits to map icons). -
Asset Pipeline: Automated promotion of AI outputs to production directories:
/assets/avatars/and/assets/maps/.
This project follows a structured, simulation-based progression to move from conceptual identity to a fully orchestrated multi-agent rescue system. Each level builds a critical technical dependency for the next.
| Level | Mission | Technical Dependency | Tech Stack |
|---|---|---|---|
| Level 0 | Identity & Baseline | Orchestration: Established the "Trinity" of user personas ( Commuter, Responder, Strategist) and the base geospatial asset pipeline. | Orchestrator/Worker Pattern, Vertex AI, Gemini 2.5 Flash, PIL |
| Level 1 | Terrain Pinpointing | Infrastructure: Implemented Model Context Protocol (MCP) for real-time vision analysis of the Mbagathi basin. | MCP, Gemini 2.5 Flash, Google Static Maps |
| Level 2 | The Pulse (SOS) | Ingest: Capturing live telemetry (SOS "Pulses") and OpenWeather data to create dynamic environment risk. | Event-driven agents, OpenWeather API, A2A communication |
| Level 3 | Graph Orchestration | Compute: Mapping the Trinity as live nodes. Calculated dynamic "Safe Edges" via GQL traversal and WKT location strings. | Cloud Spanner Graph (GQL), Google Cloud, Python |
| Level 4 | Coordinate group rescue | Orchestration: Multi-agent coordination to prevent traffic bottlenecks on "Safe Ridges" during mass evacuation events. | Agent orchestration, consensus protocols |
We leverage a hybrid stack that moves from rapid AI prototyping to high-scale cloud infrastructure.
| Environment | Purpose | Core Technologies |
|---|---|---|
| AI Studio | Prototyping | Gemma 4 31B (Multimodal Reasoning) |
| Vertex AI | Orchestration | Gemini 2.5 Flash (Multimodal Persona Consistency) |
| Kaggle | Data Engineering | Geospatial Notebooks, NASA SRTM Datasets |
| GitHub | Source & CI/CD | Python, Model Context Protocol (MCP) and FastMCP |
| Google Cloud | Production Scale | Cloud Spanner Graph(Live/Seeded), FastAPI, Cloud Run, WKT (Well-Known Text) Spatial Modeling |
🛡️ Infrastructure Resilience
To ensure the "Digital Guardian" survives unstable environments, Level 3 implemented a Dual-Redundancy DDL pattern. The system features a self-healing initialization logic that prioritizes local schema files but maintains a hardcoded DDL backup, ensuring the Mbagathi Property Graph can be reconstituted anywhere, anytime.
🛰️ 7. Technical Deep Dive: The "Hidden River" Problem
Standard maps often fail in Nairobi because the Mbagathi River is obscured by urban canopy and informal settlements. FloodPulse solves this through a "Multi-Sensor Fusion" approach:
- Multimodal Spatial Reasoning: AI analyzes soil moisture (spectral darks) to "see" the true path.
- Multi-Sector Tactical Spread: Automated generation of Zoom-17 mission tiles for distinct topographical zones (Sump, Arterial, Ridge) to ensure high-definition context for local AI reasoning.
- NASA SRTM Integration: Mathematical validation of elevation for every "Safe Ridge."
- SAR (Synthetic Aperture Radar) Capability: Integrating Sentinel-1 SAR to detect standing water through cloud cover in real-time.
- Inference Speed: < 2 seconds for local image-to-risk analysis.
- Offline Parity: 100% of core safety features must work in Airplane Mode.
- Validation: 90% alignment between AI-predicted flood zones and UNOSAT post-disaster maps.
| Agent | Landmark | Coordinates | Risk profile |
|---|---|---|---|
| Sarah - Stranded Commuter | T-Mall Underpass | -1.3148, 36.8115 |
Hydrological Low Point |
| Juma - Local/Boda Responder | Lang'ata Arterial | -1.3165, 36.8135 |
Infrastructure Bottleneck |
| Kamau - Urban Strategist | Madaraka Ridge | -1.3110, 36.8185 |
Strategic High Ground |