Manatee Civic AI: A Localized AI Governance Platform for Government Agencies

Manatee Civic AI is an AI platform specifically designed for government agencies, with the core concept of "local-first, governance-built-in" to address challenges such as privacy protection, decision transparency, and audit compliance in government AI applications. The platform includes 4 professional Agents, a complete governance layer (PII desensitization, security gating, audit logs, model registry), supports local LLM inference (Ollama air-gapped deployment) and integration with cloud models (Azure OpenAI). All data is processed locally by default to ensure citizen data security.

Government agencies face three core challenges when adopting AI technologies: citizen data privacy protection, transparency requirements for decision-making processes, and audit needs in line with public record regulations. Traditional cloud AI services are difficult to adopt by government departments due to cross-border data risks. Manatee Civic AI was created to address these pain points, with its core design concept of "local-first, governance-built-in"—all functions can run on the local network, and governance mechanisms such as PII desensitization, security gating, and audit logs are built-in.

The platform includes 4 professional Agents:

1. Civic AI Policy Agent (Ready): Provides government AI implementation frameworks, NIST RMF guidance, and 8 cases from similar counties/cities. Its knowledge base contains 13 indexed documents (over 38,000 words);
2. Citizen Service Agent (Ready): A public-facing county service chatbot that supports department routing, 311 requests, FAQ matching, and human escalation, with built-in contact information and addresses for 5 departments;
3. Web Intelligence Agent (Framework Stage): Monitors Florida AI legislation, similar county/city projects, and federal framework updates to generate leadership briefs;
4. Document Analysis Agent (Framework Stage): Performs RAG processing on county documents, automatically ingests knowledge bases, and provides searchable Q&A with citations.

Governance is the core differentiating feature of the platform, and all AI operations go through the following modules:

1. PII Desensitization: Scans and desensitizes sensitive information such as SSN and email before text reaches the LLM or is stored;
2. Security Gating: Pre-checks content safety thresholds before AI operations, detecting injection attacks and jailbreak attempts;
3. Audit Logs: Records timestamps, user IDs, and event types for all AI operations, supports multi-dimensional queries, and meets public record regulations and FOIA compliance;
4. Model Registry: Tracks model and prompt versions, and manages promotion status from development to production.

Technical Implementation: The core is an OpenAI-compatible Governed Proxy located between county applications and LLMs. Each request must go through PII desensitization, security gating, and audit log recording, supporting plug-and-play replacement of existing tools. Inference Options:

• Local LLM Gateway: Achieves air-gapped inference via Ollama, no API key required, data does not leave the machine, supports 6 county-safe models such as phi4 and llama3.1:8b;
• Azure OpenAI Integration: Supports cloud models like gpt-4o, but requests still pass through the local governance layer.

Typical application scenarios of the platform include:

1. 311 Service Desk: AI drafts responses to citizen inquiries, processes them after PII desensitization, and interaction records are logged in audit logs; 2.** Policy Drafting**: AI assists in drafting proposals, Golden Record Analyzer compares versions, and security gating blocks non-compliant prompts; 3.** Committee Meeting Preparation**: Document Analysis Agent searches county ordinances, Policy Agent provides framework guidance, and query logs ensure transparency; 4.** IT Helpdesk Classification**: AI classifies and routes issues, and the governance proxy ensures no PII reaches the LLM; 5.** Public Record Requests**: Document Analysis Agent searches indexed documents, and the audit trail supports FOIA compliance.

Manatee Civic AI provides a government AI solution that balances functionality and compliance. Through its local-first architecture, built-in governance layer, and professional Agent design, it demonstrates the possibility of AI creating value for public services under strict privacy and audit requirements. This project has important reference value for researchers and practitioners focusing on government AI applications, data privacy protection, and AI governance.