Deploying LLMs Where the Data Can Never Touch the Internet
How we stood up open-source LLMs inside a hospital network with zero internet access, and what broke first when cloud assumptions met an air gap.
Blog
Notes on building production software, machine learning, and automation systems.
Row-Level vs Schema-Per-Tenant: Two Multi-Tenant SaaS Backends, Two Different Isolation Bets
A practitioner comparison of JWT row-level tenant scoping versus PostgreSQL schema-per-tenant isolation, drawn from building both back to back in production.
Turning a Multi-Day Genomic Pipeline Into an Hours-Long One Without Losing Trust in the Output
How we cut a nanopore genomic pipeline from days to hours by building stage-by-stage QC checkpoints that catch silent failures before they reach a clinician.
The Chunking Decision That Made or Broke Retrieval in a Multi-Tenant Document RAG System
Why fixed-size chunking quietly broke retrieval across tenants, and how structure-aware chunking plus per-tenant calibration fixed it.
Extracting Trustworthy Tables From Messy Third-Party PDFs at Clinical Scale
How we turned unpredictable third-party lab PDFs into reliable clinical reports using sample-ID prediction and layered validation checkpoints.
What Focal Loss Actually Fixed in a Microscopy Image Classifier
How switching from cross-entropy to focal loss, paired with stratified 5-fold validation, fixed a microscopy classifier that architecture tuning alone couldn't.