This is a submission for the Hacktoberfest Open-Source AI Challenge Week 3: Slow and Steady
What I Built
LifeAdmin-AI is a document assistant designed to turn notices and other text documents into actionable tasks.
Instead of manually reading an entire notice and trying to remember every deadline, users can upload a PDF or TXT file, or paste text, and get a structured summary of the tasks, deadlines, priorities, and supporting evidence.
The goal is simple: make everyday administrative work easier and reduce the chance of overlooking an important deadline.
Demo
LifeAdmin-AI currently runs locally on my Windows computer using Streamlit and Ollama with the gemma3:4b model.
To try the demo, the app and Ollama need to be running on the same computer.
demo.mp4 - Google Drive
drive.google.com
Code
GitHub repository: https://github.com/kanan-pm07/LifeAdmin-AI
The repository contains the application code and setup instructions.
How I Built It
I built LifeAdmin-AI using:
Python for application logic
Streamlit for the user interface
Ollama and Gemma 3 (gemma3:4b) for local language-model inference
pypdf for extracting text from PDF documents
The application asks the model to return structured JSON containing a summary, tasks, and uncertainties. Python validation checks the response structure before the results are displayed.
I also focused on grounding extracted tasks in evidence from the original document and flagging uncertain details instead of silently treating them as facts.
Running the model locally means the application does not need a hosted model API for its current demo setup.
Why Does Open Innovation Matter?
Open-source tools make it possible to experiment with AI without depending entirely on closed, hosted APIs.
Using Ollama and an open-weight model gave me a practical way to explore local inference, structured outputs, and validation while keeping the application accessible for further development.
It also helped me understand that building an AI application is not just about generating a response. The application needs to validate that response and make uncertainty visible to the user.
What I Learned
One of the biggest challenges was getting consistent structured JSON from a language model.
I learned to combine prompt instructions with Python-side validation instead of relying on the model's output alone. I also learned the importance of testing deadlines, evidence excerpts, and ambiguous instructions with sample documents.
LifeAdmin AI is still a work in progress, and I plan to keep improving its reliability and usability.
What's Next?
Test with more types of notices and documents
Improve handling of ambiguous deadlines and conditional tasks
Make the demo easier for others to try
Continue refining validation and error handling
Thanks for checking out LifeAdmin AI!
The Lazy Payoff
LifeAdmin AI is designed to reduce the effort of reading long notices and manually extracting tasks and deadlines.
Instead of scanning a document, identifying action items, and rewriting them into a checklist, users can upload a PDF or TXT file and receive a structured summary with tasks, deadlines, priorities, and supporting evidence.
The biggest benefit is reducing repetitive reading and organizing. The user still needs to verify important dates and ambiguous instructions, but the first pass becomes much easier.
I'm currently testing the app locally and plan to measure its accuracy and time savings as I test it with more documents.
Prize Categories
Best Use of Gemma
LifeAdmin AI uses Google's open-weight Gemma 3 (gemma3:4b) model through Ollama for local document analysis. The model extracts tasks, deadlines, priorities, and supporting evidence into structured JSON, while Python validation checks the response before displaying it.
Using Gemma locally makes it possible to experiment with document automation without relying on a hosted model API for inference in the current demo setup.
Top comments (0)