Portfolio · Case study
Professional Intelligence Watch
Turning a fast-moving field into reliable, traceable professional intelligence
Professional Intelligence Watch is a governed, AI-supported knowledge system I designed to help me follow meaningful developments in AI, documentation and knowledge systems without starting from scratch every day.
It brings together ongoing research, comparison with what is already known, supporting evidence and human editorial judgement to produce a selective professional intelligence publication.
The challenge
Keeping up with developments in AI and knowledge work is not simply a matter of finding the latest articles.
Information arrives constantly from vendors, practitioners, researchers, professional communities and the wider web. Some developments are genuinely significant. Others repeat earlier announcements, add limited new evidence or present familiar ideas in different language.
Search tools and AI-generated summaries can help find and condense information. But on their own they do not reliably answer the questions I was most interested in:
- Is this actually new?
- Does it change what I already understand about the subject?
- Is the evidence strong enough to rely on?
- Is it important enough to bring to a professional reader's attention?
I wanted a more disciplined way to follow a changing field — something more useful than a daily AI newsfeed. I wanted the Watch to distinguish genuinely important developments from repetition, build on what had already been learned, and keep the evidence behind those judgements visible.
That led me to design a continuing intelligence system that could answer three questions:
What has genuinely changed? Why does it matter? What evidence supports that judgement?
My approach
I designed and implemented Professional Intelligence Watch as a continuing, AI-supported research and knowledge system.
It monitors a selected group of practitioners and sources that are particularly relevant to my field, while also looking more broadly for important developments and primary evidence.
AI supports research, comparison and synthesis. But finding something is not the same as accepting it as reliable professional intelligence.
The system therefore compares new evidence with what is already known, preserves the evidence behind important conclusions and applies explicit editorial controls before anything reaches the reader.
That distinction became one of the central design principles:
New information is not automatically new knowledge.
The Watch is designed to be selective. If nothing sufficiently important has changed, producing more content is not considered a better outcome.
How the Watch works
- 01
Discover
The Watch combines focused monitoring with wider discovery.
- 02
Compare
New findings are considered against the Watch's existing intelligence record.
- 03
Preserve evidence
Important findings retain a connection to the evidence supporting them.
- 04
Synthesise
AI assists with comparison and synthesis, bringing related evidence together and helping identify what has materially changed.
- 05
Validate & publish
The accepted reader edition is checked before it becomes the published Watch.
Read more about the five stages
1. Discover
The Watch combines focused monitoring with wider discovery.
A selected group of relevant professional sources provides continuity, while broader research helps identify developments that may sit outside that core group.
2. Compare
New findings are considered against the Watch's existing intelligence record.
That allows the system to distinguish between:
- something genuinely new;
- useful confirmation of something already understood;
- evidence that changes an earlier conclusion;
- and information that adds little enough to leave the current position unchanged.
3. Preserve the evidence
Important findings retain a connection to the evidence supporting them.
A conclusion therefore does not stand on its own. It remains possible to understand where it came from and what supports it.
4. Synthesise
AI assists with comparison and synthesis, bringing related evidence together and helping identify what has materially changed.
5. Validate and publish
Publication is a separate controlled step.
The accepted reader edition is checked before it becomes the published Watch.
The reader sees a selective professional intelligence product rather than an unfiltered stream of everything the system found.
What the reader sees
Professional Intelligence Watch is a live professional publication with a reader experience designed around continuity.
TODAY
The current successfully published edition.
RECENT EDITIONS
The ten preceding production editions, making recent developments easy to follow.
BROWSE ARCHIVE
Older editions organised by month, preserving a longer-term history.
PILOT ARCHIVE
The original pilot editions, deliberately separated from the production Watch.
Each edition links readers back to the supporting public sources behind the intelligence.
Together, those elements create a continuing professional record rather than a disconnected collection of daily summaries.
7 October 2026 · Reasoning proof
A real example
The 7 October 2026 edition shows how the Watch builds on earlier intelligence rather than treating each new source as a separate story.
Explore the reasoning behind this edition
What the Watch already knew
Previous editions had already been developing a picture around structured and governed knowledge for AI systems.
On 27 September, Meta provided production evidence that structured, auditable knowledge could operate as part of real agent infrastructure.
A later Watch then surfaced an important counter-signal: documentation agents could still miss decisive evidence, make plausible assumptions and fail to cover all affected information surfaces.
What appeared
By 7 October, a newly published empirical preprint, SciKGExtract, tested several of the controls that had previously appeared mainly as architectural ideas — including schema-guided extraction, domain normalisation and agent-based verification.
What changed
The results were useful because they were mixed.
In one evaluated task, agentic refinement improved exact-match extraction performance substantially. But a more complex case still performed poorly.
How the Watch treated it
The Watch therefore did not turn the study into a broad claim that structured knowledge or agentic verification automatically solves reliability.
Instead, it updated the existing position:
Structure helps. Normalisation helps. Verification helps. Complexity does not disappear.
That distinction matters.
The new evidence strengthened the case that knowledge quality can be engineered at several control points, while the failure case kept the conclusion appropriately bounded.
This is the kind of judgement the Watch is designed to make: not simply identifying new information, but deciding how new evidence changes — or qualifies — what was already understood.