Use this page to put your teams planning and progress reports. Project Progress Reports: * Stage I Report:
We met with our client, learned about the problem and what they wanted us to do, and gathered requirements. We tested a few different LLMS data extraction capabilities on the first basic pdf example he gave us and recorded results. Performed a lot of research to decide on our tech stack and project’s direction.
* Stage II Report:
We created an MVP just using Gemini as it was free for us students and performs well. It allows for upload of a pdf and then extracts the information it can, returning the results in JSON. We also got access to OpenRouter so we were able to start testing many more models, and we got an API key that would allow us to use any of those models in our application once we build it. We also started doing some prompt engineering in OpenRouter to test our sub-agents capabilities and prepare for our main prototype.
* Stage III Report:
We created our prototypes. The first is a data extraction library for excel that we created to extract data from diagrams created in on excel spreadsheets, as both machines and AI models cannot read that data. We also created our main prototype - the advanced multi-agent workflow. It consists of the main agent and 6 sub-agents to run individual tasks. This multi-agent workflow works on all files except excel files, hence why we must have both prototypes in order to be successful. We also created more example diagrams to test and tested our prototypes to make sure they were working.
* Stage IV Report:
We combined those prototypes into one backend service, so our main multi-agent workflow can now accept all inputs including excel. We did a lot of iterating on our final product, fixing errors we came across such as an entity aliasing error where different entities were being treated as the same entity because of similar (but unique) names. We also added another feature where users can view the output of our program in a tree-like format in the terminal to make manually checking the results of our program easier. Along with that, we added recursion so that the verifier agent now tries to fix its own errors a few times before outputting results. Finally, we also completed our documentation for our client.
Meeting Minutes: * Week1 * Week2 * Week3 etc