• 

  Visual pipeline builder—select data sources, customize AI processing, and deploy automated workflows without writing code

• 

  AI-generated weekly research digest from arXiv (John Oliver-inspired voice)—automatically updated and deployed via GitHub Pages.

• 

  Live market digest: specific crypto prices, stock indices, and AI-generated analysis updated every 6 hours

• 

  4-agent pipeline executing in sequence: Fetch → Analyze → Generate → Deploy

• 

  Each agent logs its output: Execution time, granular cost tracking, and additional details.

AgentFlow

1. Live Site: https://agentflow.neevs.io
2. Research Roundup: https://agentflow.neevs.io/artifacts/research-roundup
3. Market Analysis: https://agentflow.neevs.io/artifacts/market-pulse
4. GitHub Repo: https://github.com/jonasneves/duke-hackathon-2025

Inspiration

Information overload is overwhelming. We wanted to create a tool that lets anyone build automated AI-powered insights without needing infrastructure, coding skills, or a big budget. Set up a weekly AI research digest in 5 minutes and never touch it again.

---

What It Does

AgentFlow is a framework for building automated AI pipelines that:

• Fetch data from APIs (arXiv, Reddit, Hacker News)
• Analyze it with AI (GPT-4o-mini for summaries and categorization)
• Generate web digests automatically
• Run on GitHub's free tier (zero infrastructure costs)

The no-code builder generates production-ready code (GitHub Actions workflows, Python agents, Astro frontend) that deploys instantly.

Examples in Production:

• Academic Research Digest (John Oliver-inspired voice)
• Stock Price Monitor (tracks 8 specific stocks on 5%+ moves and publish a written report)

---

How We Built It

Tech Stack

• AI: OpenAI GPT-4o-mini (summarization, categorization, editorial writing)
• Orchestration: GitHub Actions (scheduled workflows)
• Frontend: Astro + vanilla JavaScript
• Data: JSON files in Git (no database)
• Deployment: GitHub Pages
• Architecture: Multi-agent system where agents pass data via JSON files, orchestrated by GitHub Actions on a schedule.

---

Challenges We Ran Into

• Data persistence without a database – solved by committing JSON to Git. Unconventional, but it provides built-in versioning and zero infrastructure costs.
• Time constraints – didn’t finish testing all data sources, so untested features were disabled and clearly marked as in development.

---

Accomplishments We’re Proud Of

• Runs in production
• Zero infrastructure costs or management – everything on free tiers (~$0.03/week AI only)

---

What We Learned

• GitHub Actions is an underrated AI compute platform
• GPT-4o-mini is powerful enough for $0.03/week
• Static sites can feel dynamic with scheduled builds
• It’s better to ship incomplete but stable features than broken ones
• Git can serve as a database for small-scale versioned data

---

What’s Next for AgentFlow

• Test and enable all data sources (RSS, GitHub, Twitter)
• Add more analysis types (clustering, anomaly detection)
• Build a marketplace of pre-configured pipelines
• Add webhook support for real-time triggers
• Enable multi-source pipelines (combine arXiv + Reddit + Hacker News)

Built With

• astro
• github-actions
• openai
• python