AI Tools for RWE · RWEdnesdays

RWE Coding Copilot

In development

A coding copilot that lives beside SAS, RStudio, and Python, not inside them. Give it your protocol, SAP, and codebook, and it does the thing most analysts skip under deadline pressure: it interviews you about every analytic decision the SAP leaves ambiguous before a single line of code exists.

1

Structured spec

Builds an analysis spec from your documents. You confirm it; the machine does not guess silently.

2

Decision interview

Date inclusivity, washout, censoring, grace periods: a fixed checklist of the decisions that quietly change results.

3

Annotated code

Copy-ready SAS, R, and Python where every decision-dependent line cites its decision id.

4

Audit guide

A timestamped Analytic Decision Guide from an append-only log. Your future QC reviewer says thanks.

rwe-copilot ingest my-study/ protocol.docx sap.docx codebook.xlsx
rwe-copilot confirm my-study/      # review spec.yaml first
rwe-copilot interview my-study/    # the part your SAP forgot

Bring your own key: works with any OpenAI-compatible endpoint (OpenRouter, a company gateway, local Ollama). Inputs are document text and codebook metadata only. The tool generates code; it never connects to data.

Structured Literature Reviewer

In design

Protocol-driven literature review with the boring parts done properly: explicit search strategy, screening criteria you set up front, extraction into evidence tables, and citations that are verified rather than hallucinated. The goal is a review you could hand to a methodologist without flinching.

Search you can reproduce. The query, databases, and date ranges are recorded as part of the output.
Verified references. Every citation is resolved against the source record before it appears in a table.
Evidence tables, not vibes. Endpoints, populations, designs, and limitations extracted into a structure you can sort and audit.

The lab

Things being built and broken on personal time. They graduate to this page when they survive contact with real workloads; they disappear when they do not. Current resident: an experiment pipeline that drafts and red-teams study documents (protocols, SAPs, manuscripts) with a panel of adversarial reviewer personas. More TBD, and "TBD" is doing honest work in that sentence.

The safety rules these tools live by

No patient-level data, ever.

Models see documents and metadata. Code runs where your data lives, under your governance, like it always has.

Your keys, your endpoint.

Bring your own key and route through whatever gateway your organization already trusts.

Every decision on the record.

Append-only decision logs and annotated code. If a reviewer asks "why this washout window," there is an answer with a timestamp.

The human signs the analysis.

These tools draft, structure, and check. They do not replace the analyst's judgment or their accountability.

Want early access or to argue about any of this? Get in touch →