Relay
2026
/
B2B
/
AI
/
Publishing
Overview
Relay reimagines the publishing workflow as human-AI collaboration, where technical reviews, citation analysis, IP governance and asset identification, accelerate editorial decisions and reduce downstream risk without compromising expert judgment. It has cut the publishing timeline from 3 months to 14 days and gained 100+ active users across 5 different publication types within 3 months of launch.
Team
2 AI Engineers Program Lead
2 AI Strategists halfway
Timeline
April - Nov 2026
Role
Design & AI Product Lead
Product surfaces
Word Add-in + Web App
the probelm
Content took 18 months to get published. None of it was the writing.
Once a document was written, it moved through editorial, technical, asset, and publishing reviews, as well as translation (a subsequent lifecycle not captured in this scope of work). Each specialist checked something the actual writer did not have the information, expertise, or confidence to resolve themselves. Most went into waiting: for a reviewer’s availability, for a correction from the author again, or for something to be verified again, without moving the content closer to publication.
The ask was to reduce the timeline from 18 months to 18 days. The first 2 phases of the project focused on the content getting published in English i.e. the scope of this case study, and getting the publishing timelines down from 4+ months to 10 days. Relay does that by reducing handoffs, moving responsibility to the actual owners of the content, and utilizing AI to flagging issues way upstream in the process so there are no downstream costs.
my contribution
I led design and product from discovery to adoption.
The full rundown: what I owned, directed, and built myself
The Solution
The 6 features that laid the foundation
01
Citation Detection & Verification
The agent flags missing citations, checks claims against sources, and surfaces supporting passages so authors can quickly verify and approve each link.
02
Multi-Agent Editorial Review
AI consolidates style, tone, language, and flow feedback into one review, merging overlapping flags into a single finding and suggested fix
03
Editor-Guided Author Decisions
Editors approve or override findings and record why. AI uses their rationale and relevant standards to guide authors before they repeat dismissed issues.
04
Semantic Asset Search & Rights Verification
Model flags images that need specialist review before publishing, with author overrides and an audit trail for every correction.
05
AI-Assisted Markup Review & Validation
Helps publishers inspect what AI changed and left untouched, prioritize violations, and see validation errors as they edit—before attempting a CMS upload.
The Solution: DEEP DIVE
Where AI earned its keep
Six AI-driven capabilities defined Relay, with each transferring responsibility while preserving expert judgment and human control.
01
Citation Detection & Source Matching
1a.
For each finding, the design principle is to return the evidence, not a verdict.
1b.
As task friction decreased, the need for verification increased.
So once the author confirms the source, I rallied for adding a check that compares the source text to the author’s claim. It judges whether the source supports it, partially supports it, contradicts it, or can’t be verified. Each finding carries the exact source passage, so the author can judge in seconds.
02
Consolidating a fragmented editorial review: Review the document, not the agents.
[ BEFORE ]
with severity, agent type, and agent type flagging same finding across diff severities
Separate agents checked citations, style, grammar, and flow. However, the agents often flagged the same sentence for different reasons. Showing all three would have tripled the noise. To do that, consolidation was an AI problem too, not just an interface one. So the system reconciled them into one finding with one suggested fix.
At first, I exposed each as its own tool. User testing showed how big a change that was for authors. They had one question: with every action, is my document closer to ready for my publisher? An interface built around agents buried that question under information they didn’t need and forced them to decode the system’s internals. So I consolidated everything into one Review experience.
Expand for past explorations:
Removed grouping by severity and agent type. Authors reviewed every suggestion anyway. Another hierarchy didn’t change what they did.
Moved resolved and stale findings (flags on content that had changed) to a separate tab. Keeping them in view preserved history but crowded active review with items that no longer needed attention.
It was front-loading a lot of these review steps that are usually handed to different authors and editors that brought down the review that used to take 55 days, down to a few days.
[ [ NOW: no severity, no agent type grouping (explicit UI decisions), and only one finding ] ]
03
Identification of Restrictive Images
Some images can’t be published without a separate review: a trademarked logo, patient-identifying details. Those reviews depend on specialists and take time, so they should start early. But these images were often caught late, or not at all, because spotting them depended on someone remembering to check.
When an image is uploaded, the system uses a multimodal model, to apply the SME’s own written criteria for classifying an image. If the image needs special review, the system flags it and holds it for submission.
[ Slot B · blue pop · replace ]
[ Slot A · replace ]
The model can be wrong, so authors can override it. Selecting “not an image of [x]” releases the image from the review hold. I made this deliberate so people stay in control of a judgment the AI can get wrong. Every correction is logged, which keeps the audit trail intact and shows where the criteria needs refining. The classification rules stop living in people’s heads and start improving with use.
"Did everything pass correctly through the agent? … Were there some things that the author accepted that they weren't supposed to accept?"
Feature 3
Retain editor judgment for author guidance and agent ground truth
The long-term vision called for fewer separate editing and review stages. To move toward that vision, the team proposed keeping editors out of the formal workflow in the initial few months itself, moving editors to more of an agent QA role, and banked on the agent being really good.
My stance: keep expert review in the MVP
I pushed back on doing that in the MVP and this is a capability I’d defend the hardest.
The problem: authors need support; agents need ground truth
Authors were making decisions that specialists used to handle, some on standards they’d never had to interpret. A suggestion appearing in the interface alone, no matter how strong or good, did not give the author the information to judge whether it is right. Moreover, for the system, capturing editor judgment was the first reliable ground truth we had, which meant we could measure where the agent and editors agreed and where they didn’t.
The decision: phase the handoff and explain recurring disagreements
I proposed a phased approach. Authors write and review first. Editors then review separately, approving or overriding findings and recording why. Those decisions became the context layer behind every finding. But I didn’t rally to surface editor reasoning on every finding. When an author and an editor have historically made the same call, there’s nothing to teach, and extra explanation is just noise. I reserved it for the moments they repeatedly disagree: “Editors had feedback on your last 2 dismissals of this type,” with the standard and how editors ruled on similar cases. The addition to the interface was a small interaction with a larger purpose: it made expert judgment visible at the point where responsibility was moving to the author and providing authors help only where they needed it, at the moment they were about to repeat a mistake. It provided a basis for confidence beyond simply trusting the agent, or worse, guessing what their response should be, which is not reliable ground truth for our systems to improve upon.
This turned the phased approach from a leap of faith into a measurable one: as agreement on a category rose, editors stepped back from it, and stayed in the loop where the agent was still unreliable.
Read more
[ Feature 3 — Slot A · staggered · replace ]
Impact
We shipped two product surfaces in five months and reached 100+ users in rollout.
What shipped—and what remains
Over five months, we took Relay from early prototypes to a production Word add-in and browser application. I led design through implementation, UAT, training, and ongoing feedback. The delivered experience moved important checks earlier, supporting decisions with source evidence, visible expert review, and a path to human help.
100+
users reached
Production Word add-in and browser application.
0 → 1
workflow built
Earlier verification, supported by evidence and expert judgment.
TBD
measured cycle-time results
Cycle-time and effort results were not verified at the time of writing. The source ambition was 18 months → 18 days; the Phase 1 cycle-time target was 55 → 14 days. Neither was measured performance.