Featured case study / AI systems + research engineering
The Evidence Machine.
Preserving a scientific legacy through evidence-driven AI orchestration.
My father, Dr. John A. Henning, spent his career researching hop genetics and developing cultivars for the brewing industry. I originally built his website to establish an online presence for that work. Over time, I saw the opportunity to build something more enduring: a comprehensive, accessible, properly sourced record of his scientific contributions.
What started as a website redesign became a project in research engineering: designing a process that could turn fragmented evidence into a reliable public resource.
01 / The challenge
The challenge was information, not interface.
The original site provided a useful starting point, but the record I wanted to preserve was scattered across registration papers, research publications, institutional records, interviews, trade coverage, and my father's own archive.
The sources were uneven: some were incomplete, some disagreed, and some referred to different people with similar names.
I initially experimented with simpler ways to collect and structure material, including an AI-assisted interview workflow that transformed answers into Markdown and JSON. But the problem was larger than gathering more content. It required a reliable method for discovering sources, extracting facts, reconciling uncertainty, checking claims, and publishing only what the evidence supported.
I built the original Eleventy site, then migrated it as-is to Astro in roughly a day.
Design exploration over approximately two weeks, with my father's input on voice and visual direction.
Evidence intake, drafting, verification, and integration; content sprint wrapped around 26 October.
The redesigned site was documented in final prelaunch preview.
02 / The evidence system
Separate the work. Connect the evidence.
I adapted ideas from earlier AI experimentation and adversarial review, organizing the research process around four specialist AI agents and an AI orchestrator. Each role had a defined input, output, and responsibility; human approval controlled what moved forward.
The source scout discovered and captured records. The cultivar curator extracted structured facts. The content synthesizer prepared reader-facing explanations. A separately scoped fact-checker reviewed claims against captured evidence without silently fixing drafts. The orchestrator integrated verified material through typed schemas, deterministic generators, and build checks.
Illustrative reconstruction of the workflow · Not an internal application screenshot
Evidence first. Publication second.
The key decision was to separate the messy evidence workspace from the clean publication surface. Unresolved information remained flagged. New primary documents could trigger fresh drafting and verification instead of being patched into an increasingly inconsistent record.
The work was a bespoke workflow for this site, with me as the human reviewer and decision-maker.
03 / Guardrails
Design for the ways a record can go wrong.
Source conflicts stayed visible for human decisions. A verification report recorded issues and corrections; schema checks and page-level audits checked that populated data actually appeared. The source, not a draft or a model's recollection, was the authority.
Separate responsibilities.
Research, synthesis, verification, and integration did not share unrestricted editorial authority. The fact-checker's operating instructions restricted its writing to the verification report.
Return to the evidence.
Late-arriving primary documents prompted re-drafting and re-verification. The process corrected records that had drifted, including a cultivar release year and an interview date.
Fail closed.
Build-time checks rejected internal tracker language and caught dropped or unused data fields before publication.
04 / Documented result
A more complete record, built to be checked.
By the end of the October content sprint, the project had assembled research sources into detailed cultivar profiles, a publication library, media records, and a 24-page website build. These are documented content and build results.
The cultivar rendering audit checked 267 populated fields, with zero left unrendered. The initial verification pass identified and remediated six moderate and three minor issues, with no critical findings. Subsequent targeted review caught and corrected a regression and an under-cited claim. Final checks found no internal tracker-language leaks in source or built output.
Website images show the prelaunch redesign preview captured on 2 November 2026.
05 / Reflection
Make delegation dependable.
The value of orchestration lies not only in delegating tasks, but in defining the boundaries, evidence requirements, and review processes that make delegated work dependable.
This project showed how methods first explored in experimental AI systems can become practical engineering tools. A next step may be to formalize the approach for other biographies or institutional archives through reusable subject definitions, claim tracking, and review tools. Those are future possibilities, not features of this first implementation.
For me, its original purpose remains central: preserving my father's life's work and making it accessible to the people who will benefit from it.