This week, Team Scanatics made exciting progress across all three AGIS use cases as we prepare for the upcoming AI Days! For medical, we generated synthetic prescription data and achieved promising results on our synthetic test samples. We also tested additional handwritten prescriptions to evaluate performance across different handwriting styles and document quality, helping us better understand the challenges of processing real-world medical documents. For gaming, we continued improving our content moderation pipeline and achieved promising initial detection results, with an F1 score of approximately 75% on a small set of brand categories. We also cleaned and prepared over 151,000 images for further development while working toward integrating the different components of our gaming pipeline. For e-commerce, we successfully demonstrated image-based product search, retrieving relevant products based on user queries. We are now exploring ways to combine visual product attributes with structured catalog information to make product identification and retrieval more effective. Alongside these developments, we have been refining our project poster and preparing to showcase our work.

We’re excited to present our AGIS project at HWCOE AI Days and UF AI Days on October 23 and 27! These events will be a great opportunity to share our progress, showcase our work across all three use cases, and connect with others interested in AI and its real-world applications. We look forward to representing Team Scanatics and sharing what we have accomplished so far!

This week, Team Scanatics presented our Preliminary Design Review (PDR) for the AGIS-sponsored Find Anything project, but apparently, the universe had other plans for us! We initially booked a room in Reed Hall, only to discover that the presentation screen wasn’t working. No problem, we thought, and quickly moved to Malachowsky Hall, hoping for better luck. However, at 11:55 AM, just five minutes before our presentation, we discovered that the second screen had also decided not to cooperate! At this point, we were convinced the universe simply wasn’t ready to see our presentation. But Team Scanatics wasn’t giving up! With no time to spare, we gathered around the table, opened our individual laptops, and delivered our presentation together. Despite the last-minute chaos, everyone adapted quickly, and the presentation went smoothly. Two rooms, two broken screens, but one successful PDR!

During our PDR, we presented the progress of our Find Anything project, focusing on our proposed hybrid AI architecture supporting three use cases: medical fax understanding, video game content moderation, and e-commerce semantic search. We explained our project scope, customer requirements, and engineering specifications, along with how we evaluated cloud-based, self-hosted, and hybrid approaches before selecting the hybrid design. We also walked through our functional architecture, technical performance measures, potential risks and mitigation strategies, and preliminary business case. Our coach and liaisons appreciated our progress and provided valuable feedback, particularly on improving model routing and exploring how multiple models could process the same input, either sequentially or in parallel. Another important discussion focused on refining our evaluation metrics to better reflect real-world performance rather than relying solely on numerical accuracy targets. We were also encouraged to prioritize a working prototype before introducing additional system complexity. This feedback helped us identify areas for improvement and gave us a clearer direction for the next phase of development.

Moving forward, our team will focus on incorporating the feedback, refining our architecture, and continuing to evaluate models and datasets across all three use cases. Our next major milestone is Prototype Inspection Day , where we aim to demonstrate an interactive prototype of our proposed system

After weeks of research, planning, and designing our system, Team Scanatics finally got to see our ideas in action! This week, all three AGIS use cases moved into early experimentation. For Medical, our initial document processing tests showed encouraging results in recognizing challenging features such as handwriting, faded text, and rotated text, giving us a strong starting point for testing more realistic medical forms. For Gaming, we tested our content moderation model on gaming screenshots, with promising early results in identifying potentially unsafe content while also flagging uncertain cases that may need human review. For E-Commerce, we tested our image analysis pipeline on roughly 1,240 product images, processing the entire batch in under ten minutes at a low cost while generating useful product information for our future search pipeline.

These are only our first results, so our next challenge is making sure they hold up across more diverse, real-world data. With our official PDR with AGIS coming up on October 14, we’re now using what we learned from these experiments to refine our designs and get ready to show how far Find Anything! has come.

This week, Team Scanatics laid the groundwork for our three use cases: extracting key information from medical fax documents, detecting inappropriate content in gaming screenshots, and powering semantic search for e-commerce. With all three running in parallel, our focus was on building a strong foundation for each.

A major milestone was drafting our Product Design Specifications (PDS). The document outlines our requirements, specifications, and four key performance measures: use-case coverage, budget, accuracy, and documentation. We presented it to our liaisons, who advised us to prioritize core outcomes like accuracy and scalability over nice-to-have features, and to keep our targets flexible since they may change through testing. The PDS has been sent to them by email, and we’ll revise it based on their feedback.This is a draft and is subject to feedback and change.

We also began our literature review. The team studied a research paper on potential approaches and evaluated several datasets, including options on Kaggle. For the medical use case, we generated a sample fax document to confirm it matched client expectations. Our liaisons suggested organizing fax data by quality level, from clean scans to blurry, handwritten, and degraded ones, and flagging low-confidence results for manual review. They also encouraged us to start prototyping early while HiPerGator resources are readily available.
Sources:
1)https://ijeret.org/index.php/ijeret/article/view/358/340

2) https://teams.public.onecdn.static.microsoft/evergreen-assets/safelinks/2/atp-safelinks.html

To stay connected, we set up a shared Slack channel with our coach and liaisons for quick communication between meetings.

Next week, we’ll revise the PDS, divide research and data responsibilities across the use cases, share our project roadmap with the liaisons, and build a fax template for collecting test data from other students. We’re on schedule and excited to start experimenting!