FatherP

Aura

Inspiration

When I joined UB this spring for my masters course, within the 1st week ,life felt overwhelming. After talking to my friends and class mates I realised this is a common feeling. So when we made a team for this hackathon we saw a perfect opportunity to solve this issue which is faced by students across universities and departments.

What it does

Our application is simplistic as well as powerful. It requires you to submit your schedule for the semesters(a simple pdf document) and based on that it marks your calendar for the term(deadlines and important dates).Not just that ! It asks you for a generic description of how you would like your current week to be like and do you have some commitments(eg."I want my week to be light , but also study for midterms. Also I have a doctors appointment"). Just with this simple statement it will generate you a schedule for that week !! Which you can refine with simple advices as if you are talking to an actual assistant! Also, if you are particular and know exactly what you want, we also provide an option of a questionnaire which would make it easier to arrive at a schedule which you would love <3.

How we built it

The UI and backend api's were pretty straight forward ! The magic was in the AI architecture we implemented. The pipeline goes like. Text description -> Atom-Of-Thought -> prompt(p1) + schedule(s1)-> LLM(1) -> schedule for the week(s2) -> Satisfied(Yes/No) -> if No -> (What you didn't like? -> AOT -> prompt to improve(p2) )**Chain-Of-Thought-> p1+p2+s2 -> LLM(1) -> loop till satisfied. !! To describe the above architecture in brief(The detailed explanation with flowchart is in our github repo) :- Atom of Thoughts (AOT) is our "planning assistant" that breaks the user's large, free-form goals (like "study for midterms") into small, structured "atoms" of information for the main LLM. Chain of Thought (COT) is our "refinement" process; when the user rejects a schedule, the AOT first analyzes the feedback (e.g., "study blocks are too long"). The main LLM then uses this AOT analysis in a COT process to reason about the mistake and generate a new, corrected schedule.

Challenges we ran into

The No1 challenge was to be able to allow diverse textual description to generate useful schedules. We did not want to put any sort of constraint on the user about the format of description. The user can give their description in any style and get their desired result similar to how its like talking to a real life assistant.The way to get ahead of this issue is to fine-tune our LLM to diverse textual descriptions , which we plan to do in the future stages. Integrating frontend and backend was a challenging task, we reiterated our frontend code couple of times to get to the final codebase.

Accomplishments that we're proud of

We are incredibly proud of our feedback loop where we generate schedules using just basic information from the user. This eliminates the need for the user to adhere to a particular format and can just give a low-level information of what they would like their week to look like and get a schedule which achieves that.!! We have successfully developed the frontend and backend while also integrating Google Calendar using OAuth within 24-hours of build cycle.

What we learned

We learned the power of various reasoning architectures like COT and AOT and how diverse textual descriptions can be tokenised to be used to fine-tune a diffusion model or a LLM based on use-case . And that more than the actual "text" what matters is "context"

What's next for FatherP

We plan to improve our LLM by fine-tuning it on diverse styles of textual description also we intend to add more features to aid students in their planning. Also we are planning to incorporate multi-modal emotion classification to allow students to keep their stress levels in check which would also help our model to generate better schedules for our users. We are also plan to add scheduling according the difficulty of the assignments. Future versions of Aura will be designed to better support neurodiverse students and individuals managing focus or mood-related challenges, ensuring accessible, inclusive scheduling for all learners.”