Extending Usability Testing and Tackling the NLP Engine

This week marks Sprint 9 for the MedReachAI project, and our primary focus is balancing active user testing with major architectural updates. As we push toward finalizing our MVP, we are learning firsthand how critical domain-specific user feedback is for building a reliable compliance tool.

Here is a breakdown of our progress, early testing insights, and what Collin and I are building this week.

Expanding the Testing Window
Our initial unmoderated usability testing phase brought in some fantastic early insights, but we wanted to ensure a more statistically significant sample size before locking in our final analysis. We have officially extended the testing window into a "Phase 2" for this sprint, and our professor is assisting by distributing the testing kit to a broader network of staff and students.

Our testing kit utilizes a hybrid data approach: a direct 9,999-record slice of the public National Downloadable File to simulate real-world processing, alongside smaller, purely synthetic datasets explicitly engineered to trigger edge cases like PII leaks and duplicate records.

Early Insights & UI Refinements
Even with the testing window still open, early feedback from a tester with a healthcare outreach background has already shaped our active Jira backlog. They uncovered two critical areas for improvement:

Algorithm Recalibration: The system is aggressively flagging standard, public provider data (like clinic addresses and business emails) as critical PII violations, which artificially lowers the dataset's overall health score. Collin will be adjusting the backend scrubbing logic this week to properly whitelist these public fields.

Resolving "Ghost Buttons": The tester noted that clicking action buttons like "Anonymize" or "Override" on the Data Review table provided no visual feedback. We are implementing toast notifications and immediate row state updates to resolve this UI friction.

Heatmap Readability: We also noticed an overlapping formatting issue on the Data Quality Heatmap's x-axis labels, which will be corrected this sprint with updated CSS configurations.

Building the Natural Language Query Engine

While monitoring the ongoing usability tests, we are diving into one of our heaviest full-stack features: the Natural Language Query engine. Currently, the query page is operating on a hardcoded mock dataset. To make this functional, we have split the feature into two sequential tickets:

Backend Translation Logic: Building a new Python API endpoint that integrates an LLM to translate plain English queries (e.g., "How many cardiologists are in Florida?") into executable Pandas dataframe filters against our processed dataset.

Frontend Wiring: Once the API contract is stable, we will strip out the mock timeouts in the React frontend, wire up the search bar, and dynamically map the returned JSON payload into the results table.

By separating the LLM integration from the UI wiring, we keep our sprint velocity predictable and ensure the frontend development isn't bottlenecked by the complexities of prompt engineering.

We are looking forward to closing out the usability testing later this week, compiling the final report, and pushing these new features to the live Firebase environment.