What started as "Abby's responses are slow" turned into an 18-hour deep dive that touched every layer of the AI stack — from GPU driver backends to embedding model race conditions to the fundamental question of why a 4-billion-parameter medical LLM was confidently inventing researcher names. By the end, Abby went from 15-25 second hallucinated responses to 2-5 second grounded answers backed by 167,000 vectors of medical knowledge — and we found that an 8-megabyte systemd memory lock was silently killing 25% of all GPU inference requests.

Abby can now answer "What concept sets do we have for diabetes?" and "How many patients are in our CDM?" with real data — queried live from the Parthenon PostgreSQL database at response time. Eight contextual tools give her awareness of concept sets, cohort definitions, vocabulary concepts, Achilles characterization stats, data quality results, cohort generation counts, CDM summaries, and analysis executions.

A productive Sunday on the Parthenon platform — 70 commits landed today covering four distinct themes: hardening the real-time Commons presence system, fixing a persistent CSRF authentication bug, overhauling the Grafana observability stack, and significantly expanding the medical knowledge base powering Abby, our AI research assistant.

We stood up the full Parthenon analyses pipeline end-to-end: 11 comparative effectiveness studies across 10 disease areas, 46 generated cohorts, and 26 executed analyses including R-based CohortMethod propensity score matching on populations up to 68,000 patients. Along the way, we found and fixed every null-safety bug that only surfaces when real analysis results hit the frontend.

Healthcare outcomes research lives and dies on precision. Cohort definitions that took hours to build — selecting patients with Type 2 Diabetes on metformin for 90+ days while excluding those with renal failure — represent months of clinical reasoning. Concept sets encoding specific drug ingredients, condition hierarchies, and procedure codes are the molecular vocabulary of a study. When those disappear, you don't just lose files. You lose the intellectual work of defining who your study is about.

On March 13–14, 2026, we lost all of it. Then, in an unrelated incident the next morning, we nearly lost everything again.

The Query Assistant received a ground-up redesign today — from a single 1,700-line monolith into a clean tabbed interface with two focused views, a live SQL execution modal with real-time PostgreSQL status feedback, and Solr-powered concept search built into every parameter input. This post walks through the architecture decisions, the UX patterns, and the production hardening that happened in rapid succession.

Today we transformed Abby from a capable AI assistant into an OHDSI domain expert backed by the largest curated outcomes research knowledge base we're aware of in any open-source platform. By the end of the day, Abby's ohdsi_papers ChromaDB collection held 79,070 SapBERT-embedded vectors spanning peer-reviewed research papers, the Book of OHDSI, HADES package documentation, and a decade of practitioner Q&A from the OHDSI forums.

Today we gave Parthenon's AI assistant a research library that most outcomes researchers would envy. Abby — our context-aware, privacy-preserving AI — now has 115,000+ SapBERT-embedded vectors spanning 2,258 peer-reviewed OHDSI papers, the complete Book of OHDSI, documentation from 30 HADES R packages, a decade of community forum Q&A, and 19 medical reference textbooks covering epidemiology, biostatistics, pharmacology, pathology, and clinical medicine.

This post tells the full story: why we built Abby, how the architecture works, what we harvested, what we learned about data quality in knowledge bases, and where we're headed next.

After losing app data to an accidental Docker volume wipe and spending 24 hours restoring it, we hardened the database architecture to eliminate this class of failure entirely. The Docker PostgreSQL container is no longer the source of truth for anything — the host PostgreSQL instance owns all persistent data, and automated backups run every 6 hours.

A big day on the Parthenon platform with two distinct workstreams landing in parallel: a fully-featured emoji reaction system for the collaborative messaging layer, and a comprehensive 136-test import suite designed to stress-test every data ingestion pathway the platform supports — including some deliberately nasty adversarial inputs.