Local governments sit on a mountain of public information including municipal codes, ordinances, permitting requirements, and service guides. Yet, hardly anyone can find information they want. Residents call city hall to ask questions that are already answered in a public document. Staff spend time answering the same questions over and over. The information exists; it just isn’t accessible.

Civic AI Agent is our answer. It’s a conversational assistant that lets residents ask questions and get accurate, cited answers drawn directly from official city documents with no hallucinations, no guessing, no dead-end search results.

We recently built a proof of concept for the City of Ruston. The architecture is designed to be replicated for any city, school district, county agency, or public organization.

What Is Civic AI Agent?

Civic AI Agent is a web-based chat interface built on AWS serverless architecture. Residents type a question like “What are the noise ordinance rules for construction?” or “How do I apply for a home occupation permit?” and receive a grounded answer with citations linking back to the exact section of the municipal code or ordinance that supports it.

The system is not a general-purpose chatbot. It only answers from documents you have explicitly indexed. That means every answer is traceable, auditable, and grounded in official city content.

Who Is It For?

• Cities and municipalities with municipal code, ordinances, city services, permitting
• School districts with policies, handbooks, enrollment requirements, board decisions
• County agencies with zoning regulations, health codes, public records
• Any organization with a large body of public-facing documents that residents or stakeholders need to navigate

Why an Agent? Not Just a Chatbot?

This is the most important architectural decision we made, and it shapes everything about where the product can go.

A chatbot answers questions. If you ask “How do I get a pet license?” it tells you the steps. That’s useful, but it’s still just a search engine with intelligent phrasing.

An AI agent can act. It has access to tools like APIs, forms, databases, external systems. Agents can execute multi-step workflows on your behalf. The difference looks like this:

Chatbot	Agent
“Here are the steps to apply for a pet license.”	“I’ve started your pet license application. Here’s what I need from you to complete it.”
“Your utility bill is due on the 15th.”	“I’ve scheduled your utility payment for the 14th. Here’s your confirmation.”
“The park permit form is at this link.”	“I’ve submitted your park reservation request for Saturday. You’ll receive a confirmation by email.”

Civic AI Agent is built on Amazon Bedrock Agents, which provides the orchestration layer that makes this possible. In its current form, the agent answers questions grounded in city documents. But because it is a true agent, not a chatbot, it can be extended with action groups that connect to permitting systems, payment processors, scheduling platforms, or any REST API the city exposes.

The foundation is already in place. Adding new capabilities is a matter of connecting new tools, not rebuilding the system.

In Practice: The City of Ruston

The City of Ruston proof of concept indexes the full municipal code and key pages from the city website. Residents ask questions and receive cited answers linked back to the source document. The City Clerk’s office manages the knowledge base, monitors resident interactions, and identifies documentation gaps through the admin portal — no engineering involvement required for day-to-day operations.

The proof of concept is the first deployment of the architecture. As it matures, we will share what we learn about resident usage patterns, the kinds of questions that surface documentation gaps, and how agent-based civic tools perform in a real municipal context.

Key Capabilities

Natural Language Q&A with Source Citations

Residents ask questions the way they would ask a person. The agent retrieves the most relevant document chunks, synthesizes an answer, and cites the source including a direct link to the original document.

Retrieval-Augmented Generation (RAG)

Civic AI Agent uses the RAG pattern, the industry standard for grounded AI responses for internal resources. Rather than relying on a language model’s training data, every answer is assembled from documents you control. The model cannot fabricate information that isn’t in your knowledge base. We use Bedrock Knowledge Bases, which can crawl websites or point to specific knowledge stores.

Admin Portal with Knowledge Base Management

City Clerk’s staff manage the system through a secure admin portal. They can trigger crawls of city websites, monitor indexing jobs, and review every interaction residents have had with the assistant. No engineering involvement is required for day-to-day operations. They never need to use the AWS console to manage their data.

Interaction Analytics

Every question and answer is stored in a structured database. Administrators can browse and search all interactions to identify common resident questions, discover gaps in documentation, and improve city services.

All interactions pass through Amazon Bedrock Guardrails, which filter harmful content, block prompt injection, and keep responses within the scope you’ve defined, a non-negotiable for a public-facing agent representing the city.

AWS Serverless Architecture

Civic AI Agent is built entirely on AWS managed services. There are no servers to patch, no infrastructure to capacity-plan, and no operational overhead beyond the application itself. It scales automatically from zero to thousands of concurrent users.

Infrastructure as Code with AWS CDK

Every resource in the system, Lambda functions, API Gateway stages, DynamoDB tables, Cognito user pools, CloudWatch alarms, is defined in code using AWS Cloud Development Kit (CDK). There is no manual console configuration. The entire infrastructure can be deployed from scratch with a single command.

CDK lets us write infrastructure in TypeScript, the same language the team already knows. Constructs are composable, type-safe, and version-controlled alongside the application code. When a new customer needs a deployment, the template is forked, a configuration file is updated, and `cdk deploy` stands up a complete, production-ready environment.

This approach means:

• Repeatability – every deployment is identical; no manual steps that can be forgotten or misconfigured
• Auditability – infrastructure changes go through code review, just like application code
• Teachability – new engineers can read the CDK stacks and understand exactly what is deployed and why

One Resource Per Layer — Environment Isolation Without Duplication

Rather than deploying entirely separate infrastructure stacks for dev and prod, each AWS service provides its own isolation mechanism:

• Bedrock Agents: One agent with two aliases — a draft alias routes to the in-development version (dev), a pinned alias routes to the stable promoted version (prod)
• Lambda: One function per handler with `dev` and `prod` aliases — the dev alias always tracks the latest version; the prod alias is explicitly promoted after testing
• API Gateway: One REST API with `dev` and `prod` stages — stage variables route each request to the matching Lambda alias at runtime, with no code duplication
• Amplify: One app with `dev` and `prod` branches — the branch determines the environment automatically
• DynamoDB / Cognito: Separate tables and user pools per environment — data and authentication isolation is required at these layers

One codebase. Two environments. No duplicated infrastructure stacks.

This pattern eliminates configuration drift between environments, reduces infrastructure cost, and mirrors exactly how AWS intends these services to be used.

Infrastructure changes (CDK stacks) follow the same review-and-promote pattern through AWS CodePipeline with an audit trail, approval gates, and a fully automated path from development code commit to production.

Observability Built In

Every deployment includes CloudWatch dashboards and alarms out of the box:

• API Gateway request volume, error rates, and p99 latency
• Lambda duration, error counts, and throttles
• DynamoDB system health
• AWS X-Ray distributed tracing for end-to-end request visibility

Why These Choices Matter

Every architectural decision in Civic AI Agent is defensible and teachable. We chose Amazon Bedrock Agents over building our own orchestration layer because the agent primitive is where AWS is investing. Guardrails, and knowledge base integration are first-class features, not glue code we have to maintain. We chose CDK over console configuration because infrastructure that only exists as clicks in a console can’t be code-reviewed, can’t be diffed, and can’t be reproduced from scratch — and that means your team doesn’t really own it. We chose aliases and stages over duplicated stacks because that is how AWS designed these services to support multiple environments.

For a technical buyer, this matters in two ways. First, nothing here is exotic. Any engineer fluent in AWS serverless patterns can read the CDK, understand the Bedrock Agents configuration, and extend the system. Second, every pattern in this build is one we teach in our instructor-led classes. When we hand this system to your team, they are not inheriting a black box. They are inheriting a reference implementation of the same patterns they will see in training.

AWS Well-Architected Framework Alignment

Civic AI Agent is built to the AWS Well-Architected Framework across all six pillars:

AWS Well-Architected Pillar	How We Address It
Operational Excellence	AWS CDK infrastructure as code, CI/CD via Amplify and CodePipeline, CloudWatch dashboards and alarms
Security	IAM least-privilege roles, Cognito MFA, Bedrock Guardrails, no hardcoded credentials
Reliability	Serverless and multi-AZ managed services
Performance Efficiency	Serverless auto-scaling and graceful error handling
Cost Optimization	Pay-per-request on Lambda, DynamoDB, and AI model invocation, no idle compute, shared resources across environments
Sustainability	Serverless eliminates over-provisioned capacity; shared Bedrock Agents reduces resource duplication

Deploying for a New City

The architecture is designed to be replicated. Each new customer gets their own AWS account with a full, independent deployment, not a shared multi-tenant system. That means complete data isolation, independent scaling, and no blast radius between customers. Because the entire stack is defined in CDK, standing up a new city is a matter of forking the template, updating a configuration file, and running cdk deploy. A new base environment is live in under an hour. The real work is defining your data sources and how you want your agents to work.

About Tech Reformers

Tech Reformers is an AWS Advanced Services Partner and AWS Authorized Training Partner (ATP). We design and build cloud-native solutions on AWS, and we train the developers who build and the DevOps engineers who maintain them.

Civic AI Agent is both a production product and a reference implementation, every architectural decision demonstrates the AWS patterns our instructors teach in the classroom. When we hand a project to a client’s engineering team, they receive working code and the training to own it.

If you’re a city, school district, or public agency interested in deploying Civic AI Agent, contact us. If you’re an engineering organization looking to upskill your team on AWS serverless and generative AI patterns, see our upcoming instructor-led training.

As Generative AI (GenAI) transforms organizations across industries, education, and the public sector, the security implications have become a critical consideration for organizations of all sizes. At Tech Reformers, we’ve helped numerous clients navigate the complex landscape of AI implementation with AWS solutions. One particularly valuable resource we recommend is AWS’s Generative AI Security Scoping Matrix – a comprehensive framework that provides clarity in an otherwise complex security environment.

Understanding the AI Security Challenge

The rapid adoption of GenAI technologies presents unique security challenges that extend beyond traditional security measures. While core security disciplines like identity management, data protection, and application security remain essential, GenAI introduces distinct risks that require specialized consideration.

Organizations often struggle to determine exactly what security measures they need based on how they’re using AI. This is where AWS’s Generative AI Security Scoping Matrix shines – it offers a structured approach to understanding and implementing appropriate security controls based on your specific AI implementation.

The Five Scopes of AI Implementation

The genius of AWS’s framework lies in its simplicity. It classifies GenAI implementations into five distinct scopes, representing increasing levels of ownership and control over the AI models and associated data.

Scope 1: Consumer Applications

At this level, your organization simply consumes public third-party generative AI services. Think of employees using applications like ChatGPT or Amazon’s PartyRock to generate ideas or content. You don’t own or see the training data or model, and you can’t modify them.

For example, a marketing team member might use a public AI chat application to brainstorm campaign ideas or generate draft copy. The security focus here is primarily on usage governance and data sharing policies.

Scope 2: Enterprise Applications

This scope involves using third-party enterprise applications with embedded generative AI features. Unlike consumer apps, these typically offer business relationships with vendors and enterprise-grade terms and conditions.

A common example is using enterprise productivity tools that incorporate AI features for meeting scheduling, email drafting, or document summarization. Security considerations expand to include vendor assessment and data handling agreements.

Amazon Q is the most capable generative AI–powered assistant for leveraging organizations’ internal data and accelerating software development.

Scope 3: Pre-trained Models

Moving into more technical implementation, Scope 3 involves building your own applications using existing third-party foundation models through APIs. You’re not modifying the model itself, but you are creating custom integrations.

Many of our clients operate at this level, building custom applications that leverage foundation models like Anthropic Claude through Amazon Bedrock APIs. These implementations often involve techniques like Retrieval-Augmented Generation (RAG) to enhance models with organization-specific information without changing the model itself.

Scope 4: Fine-tuned Models

At this scope, organizations take existing foundation models and fine-tune them with specific business data. This creates specialized versions of models tailored to particular domains or tasks.

For instance, a healthcare client might fine-tune a foundation model with medical terminology and documentation standards to improve summarization of patient records in an EHR system. Security considerations now extend to the training data and the resulting specialized model.

Scope 5: Self-trained Models

The most comprehensive level involves building and training generative AI models from scratch using organizational data. This offers maximum control but also requires the most extensive security considerations.

An example might be creating a custom model for specialized video content generation in the media industry. At this level, organizations own every aspect of the model development and deployment.

Security Priorities Across the AI Lifecycle

AWS’s matrix identifies five key security disciplines that span these different implementation types:

Governance and Compliance

As you move from consumer applications to self-trained models, governance requirements increase significantly. For Scopes 1 and 2, focus on terms of service compliance and establishing clear usage policies. For higher scopes, comprehensive governance frameworks become essential, including model development standards and continuous monitoring processes.

Legal and Privacy Considerations

Different regulatory requirements apply based on your implementation scope. Consumer applications require careful attention to data submission practices, while enterprise models demand robust contractual protections. For self-trained models, compliance with relevant data privacy laws becomes paramount, particularly when handling sensitive information.

Risk Management Strategies

The matrix helps identify potential threats specific to each implementation type. For instance, in Scope 3 implementations, organizations should focus on prompt injection vulnerabilities and API security. For Scopes 4 and 5, data poisoning and model extraction attacks become more significant concerns.

Security Controls Implementation

Practical security measures vary across scopes. Lower scopes emphasize access controls and data handling policies, while higher scopes require technical safeguards like training data validation, adversarial testing, and model output filtering.

Resilience Planning

Building resilient AI systems means addressing availability requirements and business continuity considerations. For mission-critical AI applications, this might include redundancy planning, fallback mechanisms, and continuous monitoring solutions.

Practical Application for Your Organization

Based on our work with clients across industries, we recommend a phased approach to implementing AWS’s security framework:

1. Accurately scope your current and planned AI implementations – Most organizations operate across multiple scopes simultaneously, so mapping your activities to the matrix is an essential first step.
2. Prioritize security disciplines based on your specific risk profile – Not all security considerations are equally important for every organization. Identify your most critical concerns based on your industry, data sensitivity, and regulatory environment.
3. Implement appropriate controls progressively – Start with fundamental security measures and expand your security program as your AI implementations mature.
4. Continuously reassess as your AI strategy evolves – The rapid advancement of AI capabilities means your security approach must be flexible and regularly updated.

Moving Forward with Confidence

The AWS Generative AI Security Scoping Matrix provides a valuable mental model for organizations navigating the complex intersection of innovation and security. By understanding where your implementations fall within this framework, you can apply appropriate security measures without unnecessarily constraining your ability to realize AI’s transformative potential.

At Tech Reformers, we’re helping organizations leverage this framework to build secure, effective AI strategies on AWS. The matrix provides clarity without oversimplification – acknowledging that different AI implementations require distinct security approaches.

To explore the complete Generative AI Security Scoping Matrix and learn more about AWS’s comprehensive approach to AI security, read more at the official AWS resource.

Need help implementing these security principles in your organization? Contact Tech Reformers today to discuss how we can help you build a secure, effective generative AI strategy on AWS.

In the ever-evolving landscape of technology, one term that’s been getting enormous attention for over a year is generative artificial intelligence (Gen AI). But what exactly is generative AI, and how can it revolutionize operations for public sector organizations? Let’s explore the transformative potential of this cutting-edge technology.

Understanding Generative AI

Gen AI refers to algorithms capable of creating new content and ideas autonomously. Whether it’s generating conversations, stories, images, or music, this technology relies on large machine learning models trained on extensive datasets. It has garnered interest for its ability to drive innovation across various market sectors, including the public sector.

Empowering Public Sector Entities

Public sector organizations face unique challenges in meeting the diverse needs of their constituents. However, AI technology presents an opportunity to address these challenges by streamlining processes, enhancing decision-making, and delivering personalized services.

Key Benefits of Gen AI

Here are some of the most compelling reasons to embrace generative AI:

• Empowering decision-makers: Government agencies can analyze vast amounts of data to identify patterns and trends, enabling proactive decision-making and efficient resource allocation.
• Accelerated research and discovery: Researchers can leverage this technology to expedite data analysis and experiment design. This facilitates faster discoveries and innovations in fields such as healthcare, education, and public policy.
• Personalized services: Generative AI enables the customization of services to individual citizens’ needs, enhancing the overall user experience and satisfaction.
• Improved productivity: Automation of routine tasks frees up human resources, allowing public sector employees to focus on big-picture and mission-critical activities and thereby achieve greater productivity.

Take the Next Step

As generative AI continues to reshape the landscape of public sector operations, now is the time to embrace this transformative technology. At Tech Reformers, we specialize in helping public sector organizations unlock the full potential of generative AI on AWS. For more information, please download this guide to getting started. Or contact us today to learn how we can help your organization thrive in the era of AI innovation.