Project Pulse
Teams remember what they decided.
Pulse captures the reason.
Role
UX Designer
Team
6 people
Time
4 weeks,
Platform
AI Hackathon
Pulse keeps a record of why each project decision was made.
Pulse drafts a record of each decision from the conversation where it was made, reasoning included. The prototype shows this in Teams. The person who made the call confirms it before anything is saved.
At a glance
The problem
Teams can find what was decided. The reasoning stays inside the call, chat or meeting where it happened, and is gone by the time someone asks.
What I owned
The research, the direction, the capture flow and the wireframe prototype, which I built with the team's UI designer.
The team
Six of us: a front-end engineer, a QA engineer, a UI designer, a colleague from HR, a mentor, and me as the only UX designer.
What shaped it
Five research findings, each with the design response it led to.
Where it ended
A working prototype covering capture, review, logs, timeline, ask and the guide. Voted the second most loved project by exhibition visitors.
Constraints
A prototype running on sample project data, with no live team to test it on.
02 · The discovery
The brief we were handed
It asked us to capture everything,
from every interaction possible.
The training team handed us one statement: never lose project knowledge again. It pointed at one answer: catch it all.
Statement 1 · as given
Never lose project knowledge again.
Project knowledge is messy and scattered. Decisions happen everywhere in Slack channels, Zoom meetings, emails, WhatsApp threads, physical sticky notes, and quick hallway chats.
Pain points it named
Managers lose track
Onboarding is slow
Traditional AI tools fail
The brief, as the training team handed it.
Every place it asked us to listen.
Research
What happens to project knowledge
between the decision and the moment someone needs it again?
I ran 8 interviews and a 14-response survey across UX, project management, engineering, QA and governance.
The evidence
7 of 8
interviewees could find the decision, but not the reasoning behind it.
The insight
People record the decision and leave out the reasoning.
What it changed in the design
The why and the option that lost became required fields on every record.
From research to design
Five findings, each with a design response.
Each row reads left to right: what the research showed, how many people raised it independently, what it told me, and the design decision it led to.
Finding
Evidence
What it told me
→
Design response
The why gap
7/8 interviews
Decisions get written down, usually without the reasoning behind them.
→
Why and the rejected alternative are required fields on every record.
Documentation loses to delivery
6/8 interviews
Any new writing task will be skipped. Even now, 5 of 8 carry decisions into Jira by hand.
→
The draft arrives written. The person checks it, fixes it and approves it.
Teams is where work happens
13/14 survey
A separate place to visit adds friction.
→
The draft lands in the Teams chat right after the call.
The need is episodic
13/14 search under an hour a week
Searching is rare, but a miss is expensive. One example: the architect's 64 tickets on an issue already solved once. Escalations and handovers are others.
→
Important decisions become project memory. Everything else stays ordinary context.
One place, many problems
7/8 interviews, split
Managers wanted one place. The architect and an auditor pushed back: more in one place doesn't mean easier to find.
→
Pulse holds the decision record. Jira, Confluence and the rest keep their artefacts, linked from it.
The evidence behind it
In their own words: why the reasoning goes missing
“
I can check the code and understand what it does, but I cannot just guess why it is created.”
Backend developer
“
We have documented the solution, but we did not document what we have tried and which didn’t work.”
Backend developer
“
It’s not laziness, but it’s not interesting.”
Senior project manager, on why documentation doesn't happen
What changed
The "why" only matters when someone questions a decision.
I reread the transcripts side by side and looked for the people who didn't fit.
Governance auditor
Found old documents easily.
Senior project manager
Was productive within a week.
Technical architect
"Putting everything in one place creates another problem."
The same moment kept coming up across four roles.
Clients raised topics that were already settled
Auditors chased escalations
The senior project manager put it to me directly:
"The frequency of the changes are humongous... how much will you track?"
Senior project manager
Example
The technical architect's team
One lost diagnosis cost a team a month.
64
tickets raised against one recurring issue
32
days spent working it out again
What happened
1
The issue was diagnosed once.
2
Nobody could find the diagnosis.
3
The team worked it out again.
With Pulse
The first diagnosis is saved as a record with its reasoning, so the next person finds it with one question.
Source
From the architect's interview. One person's account, not independently checked.
Business impact
Where the lost reasoning shows up
Repeated work
Teams re-solve problems that were already solved, like the 32-day re-diagnosis.
Cluster I · 6 people
Slower onboarding
New joiners spend their first weeks piecing together why past decisions were made.
Cluster H · 4 people
Dependency on individuals
The reasoning lives in a few people's private notes and memory.
Cluster F · 3 people
Difficult escalations and root-cause reviews
Escalations and root-cause reviews need the why most, and it often isn't there.
Cluster S · 4 people
Knowledge loss
When someone moves on, their reasoning goes with them.
Cluster O · 4 people
Client friction
Clients reopen topics that were already settled.
Cluster I · 6 people
The gap
The reasoning gets discussed in calls and chats, then rarely makes it into the documents.
What tools assume
Conversation
Decision
Documentation
What actually happens
Conversation
Lost layer
Why
Constraints
Trade-offs
Alternatives
Decision
Documentation
9 of 14
search for old project knowledge monthly or less
13 of 14
spend under an hour a week searching
14 of 14
have hit missing decision context, 9 of them sometimes
Survey of 14, self-reported
How Might We · v2
How Might We · v3
Early notes
Three versions of the question, on paper. My early notes had already circled meetings and calls as the gray area.
How I got there
8 + 14
Interviews and survey responses
Interviews across delivery, engineering and governance roles. Survey responses from individual contributors, to balance the interviews.
19
Affinity clusters
Each one a point two or more people made.
12
Themes
Related clusters merged, each counted out of 8.
5
Findings
The ones that shaped the design.
i
How to read the numbers: interview figures are out of 8 people, survey figures out of 14. The interviews leaned toward managers and governance, so the survey went to developers, designers and QA. Two samples from opposite ends, used to check each other.
Who I spoke to, why research came first, and how 8 interviews became 5 findings.
Who I interviewed
- UX Manager
- Project Manager
- Project Manager
- Technical Architect
- Full-stack Developer
- QA Engineer
- Governance Manager
- Governance Engineer
Why research came first
The team wanted to build first and work backwards from the app.
Interviews and surveys felt like time we didn't have, and they wanted to capture everything, as the brief suggested. I explained what research would tell us that the brief couldn't, and they agreed.
I ran the research.
The survey question I cut
"What features would you want in an AI knowledge platform?"
It asks people to design the product, and they answer with whatever AI they've seen lately.
The survey asks about what already goes wrong. Its one AI question asks what you'd type first.
From statements to findings
- Tag
Every statement tagged by person and number, like
KUN-4, so any claim traces back to a transcript. - Cluster
A point two or more people made became a cluster: 19 in all, each named as a plain restatement of what was said.
- Theme
Related clusters merged into 12 themes, each counted by how many of the 8 raised it on their own.
- Explain
Insights say why each pattern happens. Anything resting on one person stayed a hypothesis.
- Shape the design
The five findings that shaped the design each trace back to a theme and its count.
What people said
Our early findings
I synthesised on paper first, and every board pointed back to one search across every tool.

1
Turning point
The shift
From catching everything to catching what matters.
Where I started
Catch it everywhere, like the brief asked.
Every channel, every conversation, every voice note.
Research
The need isn't to preserve everything.
Two participants warned that tracking more has its own cost: one about over-tracking, another about notification load.
Reframed opportunity
Capture the context behind decisions that matter.
Mainly decisions that are likely to be escalated, audited or reversed.
Before · what the brief pointed at
Teams, Outlook, Jira, meetings and documents each hold a piece. None of them holds the whole decision.
After · what we kept
One record per decision that matters, with the why and the option that lost right under it. Redrawn from the SMS 2FA sample.
2
Turning point
03 · The design choice
I recommended capture first
6
Concepts after ideation
After ideation we had six concepts and limited time. Most of them needed past decisions to read from, and the research said the reasoning wasn't being kept anywhere. So I argued for building capture first, and our mentor agreed.
AI Decision Capture
Drafts each decision from the conversation, for a person to confirm.
Project Memory Timeline
The project's history as a sequence of events.
Decision Dependency Map
How each decision connects to its cause and what it affects.
Knowledge Journey Search
Ask why something was decided, and get the reasoning back.
Living Documentation
Pages that update when a linked decision changes.
Onboarding Copilot
Explains the project to a new joiner from its real history.
How it fits together
Four layers, one record
A decision-memory layer.
Across the tools a team already uses.
Messages, files and tickets stay where they are. Pulse keeps the part none of them hold: why a decision was made.
Layer 1
Existing project tools
Where work already happens. Pulse doesn't move or replace anything.
In the prototype
Teams
Concept only
Slack
Zoom
Outlook
Jira
Azure DevOps
Confluence
GitHub
Figma
Layer 2
Pulse intelligence
Detects that a decision was made and extracts the reasoning, the rejected option, the people and the source. Then a person confirms it. Connected sources are the main route (Teams in the prototype), and manual entry covers channels nothing is listening to.
Layer 3
Project memory
One confirmed decision record. The single object every view in the product reads from.
Layer 4
Project experiences
Five ways into the same memory, depending on the question being asked.
Timeline
Decision logs
Search and ask
Documents
Project guide
Where Pulse sits
Next to the tools teams already have
One judge asked why teams need one more platform. Here's where Pulse sits next to what teams already use.
Enterprise search · Glean, Rovo, M365 Copilot
Good at finding what's already written down. If nobody wrote the reasoning down, there isn't much for it to find.
Meeting assistants · Copilot recap, Fireflies, Otter
Summarise the whole call. The decision and the reason may be in there somewhere, but nobody pulls them out or checks them.
Wikis · Confluence, Notion
Hold whatever someone takes the time to write. That's the step the research says gets skipped.
Pulse
Drafts the decision, the reason and the option that lost in the Teams chat after the call. The person who made the call confirms it there. Nobody writes anything new.
Layer 2 · Pulse intelligence
AI drafts, a person confirms
Capture has to cost almost nothing,
and still be right.
Capture can fail in two opposite ways.
Too much friction
People stop responding to the prompt.
Too little
A wrong record gets saved without anyone noticing, then shows up in every other view.
Started from findings 02 and 03
What does Pulse listen to?
13 of 14
people we surveyed work in Teams, so the demo was built there.
5 of 8
interviews named manual write-up as the bottleneck, so drafting from these channels is the default.
Listened to
Teams
Shown in the prototype for Teams. Slack and other communication channels would work the same way.
Added by hand
Everything else at work
Phone calls, in-person decisions and client meetings under NDA. Paste the minutes, and Pulse drafts them.
Never listened to
WhatsApp, hallways, phones
These are personal channels. Pulse stays out of them, even when a decision happens there.
How capture was decided · 3 options considered
A template people fill in
Adds a writing task, which finding 02 says gets skipped.
Rejected
An unstructured meeting summary
Captures the conversation but not the decision, so the reasoning still has to be dug out later.
Rejected
Fully automatic capture
No friction, but nothing stops a wrong record from spreading.
Rejected
Draft first, a person confirms
The draft arrives written. The person's job is to check it, fix it and approve it.
Kept
Who decides it's right?
The person decides
The person who made the call approves it, edits it, or saves it as misc when it's useful context but not a decision. If Pulse marks a discussion as a decision, Save as Misc fixes it in one step.
Uncertain drafts are flagged
Anything Pulse isn't sure about stays marked "Needs confirmation", so an unchecked draft never looks approved.
The score is a placeholder
In the prototype, the confidence score is a fixed number on sample data. It isn't calibrated, so a high score doesn't mean the draft is right. The person confirming it decides that.
Approval and edit data could help evaluate and improve future drafts, but that needs a defined feedback loop, and the prototype doesn't have one. A score also isn't evidence. Showing the line from the call behind each why is the next step.
Where the AI can get it wrong
A wrong record is worse than a missing one, because a new joiner will trust it.
Failure
In the prototype
Low confidence
Designed: marked "Needs confirmation", with its score
Nothing was recorded
Designed: Ask says so instead of guessing
A reason nobody actually said
Not yet. The draft shows only the final wording, to keep approval quick. Next: show the line from the call behind each "why".
No clear decision, or two that conflict
Designed: if Pulse can't tell a decision from a discussion, it gives the draft a lower confidence score and leaves it unconfirmed for the person to decide. Two drafts that conflict: not yet.
04 · The experience
Months in, someone asks why.
The same question, without Pulse and with it.
Sample project data, built from the research. The SMS 2FA reason sits in the sample record. The AI didn't work it out.
Without Pulse
Ask around, then guess
A developer joins Project Orion. Her first ticket asks to bring SMS codes back for patients.
Jira shows SMS 2FA was removed. It doesn't say why.
The requirements doc describes the authenticator flow, not the option it replaced.
The reasoning was given on a Teams call. The person who made the case has moved projects.
Without the reasoning, someone rebuilds it from memory, or SMS codes come back.
With Pulse
Ask once, get the why
She asks Pulse: why was SMS 2FA dropped in Project Orion?
She answers the ticket with a link to the record.
The answer leads with the reason, with the record and its sources underneath.
How it works
Capture once. Find it when it matters.
01 · Capture
A draft appears in the Teams chat, right after the call.
When the call ends, Pulse posts its drafts in the same chat, so nobody opens another app to capture a decision.

The Teams plugin. A decision and a discussion from the same call, each waiting for a person.
01
Pulse suggests the work item the decision belongs to. Nothing is linked until someone confirms.
02
Approve and save, edit, or save as misc, without leaving Teams.
03
A discussion that wasn't a decision gets its own label, with no decision detected.
The fallback: for a call Pulse can't hear, paste the minutes or type the decision in.
The AI never saves on its own. It does the slow part, the writing, and the person keeps the final say.
02 · Review
The person who made the call checks the why and fixes what's wrong.

The review form in the web app. Every field can be edited before anything is saved.
01
Every field is editable, so fixing a draft is quicker than writing one.
02
Why and the option that lost are fields the reviewer checks on every draft.
03
Nothing is saved until a person approves it.
03 · Store
One confirmed record becomes project memory.
The decision log. Each row leads with what was decided and why, with the approver and source beside it.
The two fields that carry the idea
Why it was decided
The option that lost
These are what most tools drop.
Also on the record
Owner
Status
Source
Linked work
Confidence
What made it matter
So the decision can be trusted and traced later.
We agreed on this record before designing the screens. Because every view reads and writes the same object, the product can grow without the views drifting apart.
04 · Retrieve
The record Ask answered from also builds the timeline and the new-joiner guide.
Timeline. Decisions in sequence with the meetings and changes around them.
Project guide. A new joiner's briefing, built from what's already been captured.
05 · The outcome
What came out of it, and what it still has to prove.
Try the prototype
The capture flow, clickable
The working prototype, start to finish. Load it here, go full screen, or open it in a new tab.
Recognition
Second most loved of 50+ projects
Voted by visitors at the AI Hackathon 2026 exhibition. That was a reaction to a live demo, not usability testing: nobody has used Pulse on a real project yet.
What visitors saw: the exhibition poster.
What people responded to
Visitors loved it. The judges pushed below the surface.
We walked visitors through the problem and the working prototype, and the judges sent written feedback afterwards. They liked the design thinking, the UX, the storytelling and the business case we drew out of "context for decisions."
What landed
The AI does the capture
Visitors responded most to Pulse picking up the write-up, instead of handing operations people one more thing to fill in by hand.
Who it landed with
Managers
Managers were the most positive. In the research, they were also the people closest to the cost of lost reasoning: escalations, handovers and reviews.
What the judges pushed on
How it would work for real
Why one more platform, how it would actually be built, what it would cost to run, whether review adds overhead, responsible AI, keeping data inside the company's own environment, and more use cases than one sample project.
Still open
What still needs testing
What's still open falls into two groups: gaps in the design, and things the research can't show yet.
Open in the design
Drafts nobody confirms
Unconfirmed drafts can become a new backlog. They need an owner fallback, a timeout and a digest.
Why the draft was flagged
The draft doesn't yet show what made it a decision, so a false positive is slow to dismiss.
Consent and retention
If Pulse drafted from real Teams, Slack or Zoom calls, it would need visible consent, permissions and retention rules. Governance will ask for those first.
Reversing a decision
Superseded is a status, but the flow that links an old decision to its replacement isn't designed yet.
Open in the evidence
No live use
Nobody has used Pulse on a real project yet.
Missing voices
We never spoke to a client, a business analyst, or someone in their first week on an established project.
Self-reported
Everything came from what people told us. Low search time could also mean people have given up looking.
What I'd test next
A new joiner and a BA, in their first week, because they're the two people Pulse most claims to help.
Teams keep the decision.
Pulse keeps the reasoning behind it.
Project Pulse
UX Designer. AI Hackathon 2026, four weeks. Research, synthesis, knowledge capture, and the wireframe prototype, built with a UI designer.
Evidence
8 interviews, 14 survey responses. Screens are wireframes from the shipped prototype. Diagrams marked Redrawn were made for this page.
