Hey everyone, Scott here. Collin and I are back with the latest engineering update for MedReach-AI.
We just officially closed out Phase A of development. Over the last two weeks, we built a rock-solid foundation for our file ingestion pipelines, set up a heuristic engine that automatically maps healthcare columns, and locked down our export endpoints with strict HIPAA-compliant filters.
With the plumbing working flawlessly, we are now shifting into Phase B: Data Intelligence. This is a massive two-week sprint where the workload essentially doubles, and we transition from simply moving data around to actively analyzing and sanitizing it.
The Sprint 3 Goal
To put it simply: we are building the brain. Our goal for Sprint 3 is to take raw, uploaded healthcare datasets and run them through automated PII/PHI scrubbing, real-time government registry validations, and machine learning anomaly detection before the user ever sees the final dashboard.
How We Are Tackling Phase B
Looking at our Jira board, we have broken this phase down into a few core pillars:
Machine Learning (The AI in MedReach-AI): This is the part I am most excited to dive into. We don't have months to train a massive neural network, so we are utilizing unsupervised machine learning via Scikit-Learn. By deploying IsolationForest and DBSCAN clustering algorithms, our backend will be able to mathematically evaluate a dataset on the fly and flag statistical outliers or corrupted records without relying on rigid, hard-coded rules.
Security & Compliance: We are integrating the Microsoft Presidio engine to actively detect and anonymize protected health information (PHI). We are also building ingestion engines for CMS Open Payments data to flag financial conflicts of interest (Sunshine Act compliance).
External Validation: We are building an asynchronous client to query the official CMS NPI Registry. This will validate provider credentials in real-time and enrich our database with verified classification statuses.
Dashboards & Deduplication: While I handle the machine learning and deduplication algorithms on the backend, Collin is building out the interactive React tabs. These interfaces will allow compliance officers to visually review our AI’s flags, merge duplicate records, and download legally sound PDF audit trails.
What’s Next?
It is a heavy workload, but our Phase A architecture is clean, and we are ready to execute. For the first week of this sprint, I am focusing entirely on getting that Scikit-Learn Isolation Forest up and running to start flagging our test data.
We will be back next week with a mid-sprint update to share how the anomaly models are performing!
Top comments (0)