Bringing routine biostatistics programming in-house with AI

Outlook emails become a searchable knowledge base by dragging them into a synced folder — one paste-able PowerShell script. No Power Automate designer, no SharePoint clicking, no Copilot Studio.

The study KB as a navigable linked-note graph, now a runnable R Shiny app (igraph + visNetwork) + a visual guide + a shareable interactive graph + a build-it walkthrough — the three CSVs are the vault. Plus the 28-note vault and design note.

The agentic trial-operations layer on the hybrid stack — recurrent QC, SOP RAG, eTMF filing, RBQM.

The local small-model trial-ops stack and the Copilot-free Study Operations knowledge base.

Smartsheet updates itself from scheduled SAS/R — both the AI-assisted and the no-code/no-AI versions.

Run a UAT of the randomization & supply (IRT/RTSM) system before go-live, from the biostatistics lens. Five ways from one shared R logic core: a visual guide (Mermaid diagrams + a live go-live gate computed from the data), a runnable R Shiny app you edit and own, a shareable dashboard, a fillable Excel workbook for the eTMF, and a build-it walkthrough. Synthetic test data.

Where do the hours come from when out-of-scope work lands? Five ways from one shared R logic core: a visual guide (decision-tree + source-map diagrams, a live net-findable roll-up), a runnable R Shiny app, a shareable dashboard, a fillable Excel workbook, and a build-it walkthrough. Net findable = (Σ planned + contingency) − Σ EAC; hours only, no dollars.

Deterministic, no-AI monitoring that detects, packages, and pages humans. The coverage playbook covers five scenarios (planned leave, unplanned absence, role handoff or departure, onboarding, bus-factor review), and the interactive Continuity Toolkit builds a coverage matrix, risk register, checklists and handoff packet in the browser.

ADaM-based Table/Listing/Figure pseudocode for every early-phase design, in both SAS and R (pharmaverse).

Quarto + on-prem MedGemma drafts the narrative text; the model never sources a number, and every draft is number-validated and human-gated.

The complete Phase-1 clin-pharm deliverables pipeline (136 deliverables) with a worked PK example.

A public-calibrated synthetic study running all the tools end-to-end (8 ADaM domains, NCA, TLFs, SLM).

Four independent Streamlit apps for clinical eCRF extracts, each a four-tier chain with its own audit log: MedDRA-style auto-coding of adverse-event verbatims, PII/PHI flagging and redaction, participant safety and discontinuation risk scoring, and anomaly and drift detection. Labeled [SI] stand-ins run until trained runtimes are installed. The public demos run in a locked hosted mode on synthetic data only (99-prefixed placeholder codes, not MedDRA), with a 244-check QC harness in CI.

Seven named gates (reproducibility, numeric-provenance, CDISC conformance, SAS↔R parity, adversarial QC, regulatory standards, AI-use governance) that make a deliverable delivery-confident. Runnable + dogfooded on this repo's own libraries + synthetic data.

The operational study-lifecycle console (portfolio RAG, ranked early-warning signals, milestones, TLF funnel), now a runnable R Shiny app you edit and own + a visual guide + shareable dashboard + build-it walkthrough, all from one shared R logic core. Plus the operating-depth field guide.

The participant-safety (TRIALMON) and trial-termination risk dashboards — the safety complement to the lifecycle console.

The biostatistics-AI training walkthrough.

Every wiki, pitch, and worked example has a narrated walkthrough (true headless-Chrome screencasts for the live tools). A few flagship videos: