Innovation Forum × R1X
Apply
Menu
Hands typing code on a laptop

Insights · Applied AI research

What the machine can read, and what it must never see

Research from the collaboration, built around worked examples rather than positions: automating invoice processing so an agent can read your figures without ever reading your business, and a governed local pipeline that turns public proceedings into podcast episodes without the source ever leaving the machine.

Featured · Build patternsInvoice automation — build patternsThe engineering notes behind the two build stages in the retrieval and invoice architecture briefing. Foundational steps are F1–F3; advanced are A1–A6.August 2026 · 12 min read · R1X engineeringRead the piece
  1. Overview · Applied AIThe public record nobody watchesThe House of Lords publishes hours of proceedings to YouTube every week. Almost nobody watches them. This is the machinery that turns that record into something a person will listen to on the way to work — built locally, governed by a spreadsheet, and lawful by construction.6 min read
  2. Architecture · Local-firstLocal-first, and what it costs youOne constraint chose this entire stack: no source text leaves the machine. That single rule rules out most of the obvious tooling, and buys something you cannot get any other way — the ability to demonstrate, rather than assert, where the data went.7 min read
  3. Governance · DataThe run book: Excel as a governed pipelineMost content pipelines are governed by a database, a ticket system, or nothing. This one is governed by an Excel workbook — deliberately, and for reasons that survive contact with an auditor.7 min read
  4. Pipeline · Data engineeringTen steps from recording to episodeThe whole flow is one command. Inside it is a deliberate, auditable ten-step journey in which every step maps to a checkbox in the run book and produces an artefact you can point at.8 min read
  5. Governance · EditorialEditorial controls: the human gate behind every episodeGenerative models are useful precisely because they are fluent. That fluency is also the risk — so editorial control is a pipeline stage here, not a hope.6 min read
  6. Data engineering · MedallionBronze, silver, gold: layering an audit trailThe pipeline is, at heart, an extract-transform-load system with an audit trail. Getting it past "produce ten episodes" means exposing that data in layers other people can consume without re-reading raw transcripts.7 min read
  7. Data science · EvaluationScoring retrieval honestlyA retrieval system is only as good as the chunks it fetches and the answer it writes from them. Without a repeatable way to score both, "improve the model" is guesswork — so this project publishes its baseline, including the parts that are bad.6 min read

Briefings

The same material, in three lengths

Every piece exists as a deck as well as an article, so it can be presented as easily as it is read.

  1. 1 min

    LLMs and retrieval in a minute

    The memorable narrative, a jargon translation table, and a sixty-second read-aloud script. Eight slides.

    One idea, no architecture

  2. 10 min

    Retrieval and invoice architecture

    The crude five-box sketch with its failures named, then the same flow engineered — plus a permissions table. Ten slides.

    The same five steps, twice

  3. 60 min

    Operational BPMN sandbox

    The running process as two pools and five lanes, what the authorisation sandbox is, and a lane-by-lane readiness table.

    Two pools, five lanes, one sandbox

Want this work applied to your problem?

Sponsors bring a real problem class to a challenge day and get a cohort working on it, with the research behind them.

Get in touch Apply for a place

Or email info@inno-forum.co.uk