Inspiration
In cities like Johannesburg, infrastructure failure isn't just an inconvenience; it's a crisis of dignity. A burst pipe in Soweto often goes unreported or sits in a backlog for weeks because the data pipeline between the Citizen and the City is broken. Current municipal systems are reactive—waiting for complaints.
We asked: What if a City didn't just listen, but actually thought? We wanted to move beyond the concept of a "Smart City" (which just watches) to an "Agentic City" (which acts). We built CivicNerve to restore trust by turning passive reports into immediate, auditable action.
What it does
CivicNerve is an autonomous municipal operating system designed to be deployed on IBM watsonx Orchestrate. It replaces the manual "call center" layer with a multi-agent AI workforce.
Citizen Reporting (The Edge): A user snaps a photo of a pothole or leak via our mobile web app.
Sensory Cortex: The AI analyzes the image for severity (e.g., distinguishing a "Major Flood" from a "Minor Drip").
Bylaw Compliance (The Brain): It uses IBM Docling to parse unstructured PDF municipal bylaws, locating the exact regulation and tariff code required for the repair.
Governance Audit (The Conscience): The Granite Guardian agent checks the estimated repair cost against the live budget. If it's safe and funded, it cryptographically signs the order.
Zero-Touch Dispatch: The work order is sent to the nearest crew instantly, appearing on the City Dashboard without human manual entry.
How we built it
We utilized the "Mellea" Generative Computing Pattern to build a Hybrid AI kernel that interleaves Creativity with Control.
The Brain: We used IBM Granite-13b (via watsonx.ai) for reasoning and intent classification.
The Ingestion Layer: We used IBM Docling to ingest complex PDF documents (like the Water Services Act) and convert them into structured decision tables.
The Governance Layer: We built a deterministic Python agent ("The Guardian") that acts as a firewall between the LLM and the real world. It audits every JSON output for safety (ISO-9001) and budget limits before minting a SHA-256 digital seal.
The Infrastructure: The entire system is containerized in Docker and exposes strict Pydantic models, making it compliant with the Model Context Protocol (MCP) for easy integration into enterprise workflows.
Challenges we ran into
The "Hallucination" Risk: Early versions of the AI would approve R10 million repairs for small leaks. We solved this by implementing "Agentic Non-Repudiation"—forcing the AI to pass a deterministic code-based audit before execution.
Unstructured Data: Municipal laws are messy. Standard RAG tools failed to read the "Fines & Tariffs" tables correctly from PDFs. IBM Docling was a game-changer here, allowing us to preserve the table structure for accurate financial calculations.
Docker Networking: Orchestrating the communication between the Streamlit containers and the FastAPI brain required complex networking logic to ensure the "Live Map" updated in real-time.
Accomplishments that we're proud of
Building a Full-Loop Agent: We didn't just build a chatbot; we built a system that takes an image and outputs a verified, dispatched work order.
The "Digital Seal": Successfully implementing a cryptographic audit layer that proves AI can be trusted with financial decisions.
Zero-Touch Latency: Achieving a workflow where a legitimate report can go from "Phone" to "Foreman" in under 10 seconds.
What we learned
Governance is greater than Intelligence: In the public sector, it doesn't matter how smart the AI is if it isn't compliant. Building the "Guardian" taught us that AI needs boundaries to be useful.
Structure Matters: Data in the real world is messy. Tools like IBM Docling are essential because they turn "documents" into "databases" that agents can actually use.
Trust is a UX Feature: Showing the user exactly what the AI is doing (Scanning -> Planning -> Auditing) via the status bar significantly increased user trust in our testing.
What's next for CivicNerve
IoT Integration: Connecting the "Sensory Cortex" directly to smart water meters for predictive maintenance before a citizen even reports it.
Multi-Modal Expansion: Adding voice support for citizens to report issues in local languages (Zulu/Xhosa) to increase accessibility.
Watsonx Deployment: Finalizing the packaging as a reusable "Custom Skill" in the IBM watsonx Orchestrate catalog for other municipalities to adopt globally.
Log in or sign up for Devpost to join the conversation.