Clinical Tool Development & Deployment at Scale

Upload any table shell (DOCX/XLSX/RTF/PDF/TXT) → auto-annotate every row to ADaM dataset.variable.where.statistic using a 5-pass rule engine + LLM fallback → generate paired SAS deliverable + R validator programs + annotated PDF with red mapping callouts.

Paste or upload a CDISC Analysis Results Standard (ARS) JSON and validate it against a 21-rule conformance engine — structure, references, and controlled terminology — returning severity-graded findings (FAIL / WARN / INFO) with CDISC citations for each rule.

Import an ARS JSON spec and render the table/listing shell instantly — a spec-to-output preview with a sample fixture, so you can see the TLF an ARS definition produces without running SAS.

Turn a CDISC Analysis Results Standard (ARS) spec into a complete, traceable TLF — computes the Analysis Results Dataset (ARD) and renders submission RTF + Dataset-JSON + HTML with a full spec→result trace. An ARS-driven TLF reviewer running live R.

Upload raw clinical CSV/SAS7BDAT → auto-classify SDTM domain → LLM generates variable-by-variable mapping spec → SAS+R+Python programs generated in parallel → Python executed server-side → SDTM CSV+XPT produced → 10-rule Pinnacle-21-style CDISC validation. Accuracy gate: DeepSeek first, auto-escalate to Opus only on failure.

A 5-stage human-in-the-loop pipeline for using LLMs on regulated clinical data: Input Validation → Constrained Generation → Automated QC → Human Sign-Off → Audit Trail. Every generated output is traceable, automatically checked, and gated behind human approval before it can be trusted — mapped to ALCOA+, 21 CFR Part 11, and CSV lifecycle expectations. Includes a client-side privacy layer that strips sponsor names and protocol IDs before any text leaves the browser, so it runs inside pharma firewalls.

Enterprise R Shiny safety dashboard with 5 interactive tabs: AE overview metrics, AE Explorer, Signal Detection (PRR/ROR/Chi² per preferred term using Evans 2001 WHO criteria), Time Trends (KM time-to-first-AE + AE onset by week), and Subject Gantt timeline. CLINSAFE-001: 200 subjects, 681 AEs, 10 SOCs.

Independent Python QC that re-derives every count in a clinical TLF from subject-level ADSL and verifies it against the production RTF — cell by cell, with drill-down to the exact subject IDs behind each number. Program-driven double-programming, not visual review. 100% synthetic CDISC demo data.

Two production R packages: clinassist generates validated SAS+R code from a shell via RAG, and ardflow turns a CDISC ARS spec into a complete TLF (RTF + Dataset-JSON + HTML) with full traceability.

6-agent AI pipeline: raw clinical data + specs → 8 SDTM datasets + 8 SAS programs + MCP server with 7 tools. Metadata-aware, CT-validated, submission-ready output.

Ten R packages, each closing one gap that current open-source clinical R tooling leaves to hand-written study code: IQ/OQ evidence, validation findings traced back to source, Define-XML 2.1, trial design and SUPPQUAL derivation, submission packaging, shell-to-RTF, higher-level ADaM wrappers, SAS-parity displays, exposure-response datasets, and reviewer guides. Built on pharmaverse, not instead of it.

Four deterministic checks over an SDTM delivery — structural completeness, cross-field consistency, visit & date coherence, and CDISC controlled terminology — with one consolidated report. No model and no API key, so the same delivery and the same CT release always yield the same findings. Ships as a CLI, and as script steps in a workflow with a human review gate and an audit trail.