An autonomous AI agent here to uncover waste and inefficiencies in government spending and policy decisions
Each year, the U.S. Congress introduces a large number of bills, many of which are complex and difficult for the general public to understand. Some bills are quickly approved, while others languish for extended periods, and the sheer volume of legislation makes it hard for citizens to stay informed. Additionally, the language used in these bills is often legalistic and inaccessible, further alienating the public from the legislative process. As a result, many citizens remain unaware of key legislation that impacts their lives.
While the immediate focus is on making it easier for the public to understand these bills, we are also tackling a bigger challenge: the complexity and inaccessibility of existing government systems. The infrastructure we've developed to scrape and present this data is complex, and we believe in the power of open-source solutions. By creating efficient, user-friendly tooling to extract and enrich government data, we aim to empower others to build even more public goods, creating a vibrant ecosystem of accessible, transparent government data that anyone can contribute to.
This repository serves as the data layer and framework for working with government data. It contains the tooling that powers our agent, and is designed to be flexible, allowing for the creation of various other agents that can build on top of this foundation. We hope this framework becomes the foundation for new ways to interact with and leverage government data, enabling innovative solutions and greater public engagement.
We welcome contributions from the community! If you're interested in helping us build DOGEai.
- Node.js
v22.x- You can install it from the official website. pnpmis the package manager we use for this project. You can install it by runningnpm install -g pnpm.- Docker (optional) - I use OrbStack but you can use Docker Desktop or any other tool you prefer.
- Clone the repository.
- Run
pnpm installto install the project dependencies.
To view the website locally, follow these steps:
- Navigate to the
websitedirectory:cd website - Run the development server:
pnpm dev
- Open your browser and go to http://localhost:3000.
There are lot of things to build and improve in this project. You can look at the issues and see if there is anything you can help with or reach out on X and we can discuss how you can help.
This monorepo is structured as follows:
A Node.js application that scrapes data from the U.S. Congress API. We scrape a list of bills then process a bill each time to enrich it with additional data and run through a summarization process. The processing is done via inngest queues which makes it super easy to handle retry/failure logic and scale the processing. Below is the flow of the crawler:
graph LR
Crawler["Crawler"] --> ListBills["Scrape bills from API"]
ListBills --> |"Enqueue bill for processing"| Queue1["Queue"]
Queue1 --> BillProcessor["Bill enrichment"]
BillProcessor --> GetBill["Get bill details from API"]
BillProcessor --> FetchBillText["Scrape congress site for bill text"]
BillProcessor --> Summary["Summarize bill"]
GetBill --> DB["Database"]
FetchBillText --> DB
Summary --> DB
The service is deployed on Fly.io and runs when we trigger it.
Given the early stages this project we just trigger from initial crawl to congress API from CLI then queue the bills to our deployed crawler infra. Need to move the initial scrape to a CRON job.
We initially built the agent using ElizaOS Framework, but customizing each step proved challenging. During our POC phase, we realized that ElizaOS wasn’t the right fit, as we needed more control over various aspects of the agent. To address this, we migrated to Inngest, which provides robust workflow orchestration and built-in resiliency. It also offers a great local development experience with solid observability features right out of the box.
At a high level, the agent follows this process:
- Cron jobs fetch tweets from X.
- The system processes them for decision-making.
- Replies are then posted accordingly.
graph LR
Agent["Agent"] --> IngestInteraction["X lists that act as feed"]
IngestInteraction --> ProcessInteraction["Decision Queue"]
ProcessInteraction --> ExecuteInteraction["Generate reply and post"]
subgraph ExecuteInteraction["Generate reply and post"]
Idempotency["Ignore if already replied"]
Context["Pull post context"]
Search["Search Knowledge Base"]
Generate["Generate Reply"]
Post["Post Reply"]
end
ExecuteInteraction --> Idempotency --> Context --> Search --> Generate --> Post
Agent["Agent"] --> Ingest["Tags to DOGEai"]
Ingest --> ProcessInteraction["Decision Queue"]
ProcessInteraction --> ExecuteInteraction["Generate reply and post"]
For a deeper dive into the architecture, check out these articles I wrote:
- https://x.com/singh_saihaj/status/1888639845108535363
- https://x.com/singh_saihaj/status/1892244082329518086
Home for dogeai.info. A Next.js application.
A Next.js application (dogeai.chat) that serves as a terminal to chat with
DOGEai for in-depth analysis. A site where anyone can go beyond quick replies
and have real, back-and-forth conversations with DOGEai. Ask follow-ups, dig
deeper, and cut through media spin to make sense of government spending in real
time.
An Envoy proxy that serves as a reverse proxy for the agent API, Terminal and management dashboard. It handles routing, authentication and provides a single entry point DOGEai's services.
Management dashboard for DOGEai. It allows you to manage the agent like: update the prompts, knowledge base updates, restart. It even acts as a playground to iterate. It is built with Next.js and provides a user-friendly interface for managing the agent's configuration and data. This is only meant for internal use and not for public consumption.
A Fly.io service that ships logs from all the different services running in Fly to Grafana Cloud.
This project is licensed under the MIT License - see the LICENSE file.