This is my submission for the MLH x DEV Writing Challenge.
What I Built
Introducing ScoutFlow AI — From Opportunity to Action.
Students and developers discover exciting opportunities every day, including hackathons, coding competitions, open-source programs, grants, workshops, and learning opportunities.
But there is a problem: information is scattered everywhere.
Important event details are hidden inside social media posts, Discord announcements, Telegram channels, posters, and lengthy PDF documents. Keeping track of registration deadlines, eligibility requirements, judging criteria, and preparation tasks becomes difficult.
That's why I built ScoutFlow AI.
ScoutFlow AI is a multimodal AI-powered opportunity intelligence and productivity platform that transforms unstructured event information into verified, actionable preparation plans.
Instead of manually reading event documents and creating tasks, users can upload a poster, screenshot, or PDF and let AI extract the important information.
✨ Key Features
1. Gemma 4 Multimodal Scanner
Upload hackathon posters, screenshots, and rulebook PDFs. The AI analyzes the content and extracts event names, deadlines, eligibility requirements, prize tracks, and technology requirements.
2. Verifiable Source Evidence
ScoutFlow doesn't just extract information. It also provides supporting quotes from source documents, helping users distinguish verified facts from assumptions.
3. Opportunity Compatibility Intelligence
Evaluates opportunities against the user's technical skills, interests, and weekly availability to provide skill-match insights and workload feasibility analysis.
4. Automated Preparation Roadmaps
Converts event requirements into structured four-phase preparation plans with time estimates and automatically organized tasks in the Kanban planner.
5. Deadline Radar
Detects overlapping deadlines within 48 hours and identifies potential workload conflicts when preparation requirements exceed available weekly hours.
6. Master Deadline Calendar
Provides a monthly calendar view of event kickoffs, submission deadlines, and preparation milestones.
7. Contextual AI Assistant
Allows users to ask strategic questions about specific opportunities, tracks, and deliverables while distinguishing stored facts from recommendations.
8. Secure Personal Workspace
Includes JWT authentication, bcrypt password hashing, private user workspaces, and strict user data isolation.
💡 The Core Workflow
ScoutFlow follows a simple process:
CAPTURE → UNDERSTAND → EVALUATE → PLAN → EXECUTE → SUBMIT
The goal isn't to build another generic chatbot or task manager. It's to help students and developers turn scattered opportunity information into meaningful action.
Demo
Explore the project:
🌐 GitHub Repository:
https://github.com/ayushkumarak018-stack/ScoutFlow-AI
What to explore
- Upload a hackathon poster or event PDF.
- Explore multimodal event extraction.
- Review source evidence and extracted information.
- Check opportunity compatibility.
- Generate a preparation roadmap.
- Explore the Kanban planner and Deadline Radar.
- View your upcoming events in the master calendar.
Partner Technologies
Google DeepMind — Gemma 4
Gemma 4 is at the center of ScoutFlow's AI architecture.
I wanted to explore how multimodal AI could solve an everyday problem rather than simply creating another chatbot.
The interesting challenge was making AI understand unstructured event information and transform it into useful, structured data.
ScoutFlow integrates Google's GenAI SDK through a modular AI provider architecture.
The intended primary Gemma-powered workflow focuses on:
- Understanding event posters and visual information.
- Interpreting screenshots and rulebook PDFs.
- Extracting structured event information.
- Identifying eligibility requirements and judging criteria.
- Generating opportunity compatibility insights.
- Creating personalized preparation roadmaps.
- Answering contextual questions about event requirements.
A particularly important part of the implementation is source verification.
AI-generated information can sometimes be incorrect, especially when interpreting dates and competition rules. ScoutFlow therefore includes source evidence and confidence information to help users review extracted details before relying on them.
Open-Weight AI Architecture
Another important design decision was making ScoutFlow flexible rather than dependent on one AI provider.
I implemented a modular provider architecture with three modes:
1. Gemma Provider
Uses Google's GenAI SDK for the primary AI integration.
2. Local Open-Weight Provider
Supports configurable OpenAI-compatible inference endpoints through tools such as Ollama and vLLM, allowing users to explore self-hosted models.
3. Demo Provider
Provides a clearly labeled demonstration mode with simulated extraction and compatibility results, without requiring an API key.
ScoutFlow explicitly identifies the active provider and model status. It does not disguise simulated results as actual model responses.
Technology Stack
Frontend
- React 18
- TypeScript
- Vite
- Tailwind CSS
- Lucide React
- Framer Motion
Backend
- Python 3.11+
- FastAPI
- SQLAlchemy 2.0
- Pydantic v2
- Bcrypt
- Python-Jose JWT authentication
AI & Document Processing
- Google GenAI SDK
- Gemma-oriented AI provider architecture
- Ollama / vLLM integration
- PyPDF2
- Pillow
Database
- SQLite for local development
- PostgreSQL for production
Testing
ScoutFlow includes automated integration tests covering important application workflows:
- User registration and JWT login.
- Multimodal scanner document extraction.
- Profile compatibility engine.
- Preparation roadmap generation.
- Automatic task creation.
- Deadline Radar conflict detection.
- Contextual AI question answering.
- Calendar event aggregation.
The project is also structured to support Docker Compose deployment.
Hackathon Experience
Participating in Hacktoberfest Hack Day 2026 gave me an opportunity to experiment with open-weight AI and explore how multimodal models can be used to solve practical problems.
While developing ScoutFlow, I wanted to go beyond the traditional chatbot experience.
I focused on building a complete workflow where an AI model could understand visual information, extract useful details, evaluate opportunities, and help users organize their next steps.
One of the most interesting parts was thinking about AI reliability. Extracting information is only one part of the problem; helping users understand where that information came from is equally important.
Working on this project also helped me explore full-stack development, API architecture, AI provider abstraction, document processing, authentication, and automated testing.
What's Next?
I want to continue improving ScoutFlow with:
- Better multimodal document understanding.
- More local open-weight model support.
- Improved team collaboration.
- Smarter opportunity recommendations.
- More personalized preparation planning.
My long-term vision is to make ScoutFlow a useful companion for students and developers who want to discover opportunities, stay organized, and turn their ideas into real projects.
A big thank you to MLH, DEV, Google DeepMind, and the Hacktoberfest community for encouraging developers to experiment, build, and learn! 💜
Built with curiosity. Powered by open innovation. Designed to turn opportunities into action.
Top comments (0)