CASE STUDY · B2B PRODUCT DESIGN
JARVIS Alerts
A false alert flow for Staqu’s AI CCTV platform that makes every correction visible, so the people who get woken up at 3 AM can finally see the system learning.
ROLE
Solo. Problem framing, flow, UI and copy
TYPE
Independent B2B concept
TIMELINE
1 day, March 2026
TOOLS
FigJam, Figma
Independent concept work. Not affiliated with Staqu. JARVIS is Staqu’s video analytics product. Camera stills, times and counts in the screens are illustrative.
THE HARDEST CALL
The false alert wasn’t the problem. The silent dismiss was.
The obvious brief is “fewer false alerts”. But that’s a model problem, and no screen I design retrains a model.
So I looked at what happens after a false alert lands. Someone opens it, sees their own car or their own night guard, and swipes it away. And then nothing. No sign the swipe counted. Tomorrow the same camera fires the same alert, and the system looks exactly as dumb as yesterday, even if it’s quietly getting better.
That’s where I spent the design. Not on another filter screen, but on making every correction visible.
A system can be improving and still lose your trust, if you never see it improve.
WHO THIS IS FOR
The person holding the phone at 3:14 AM
JARVIS turns a business’s existing CCTV into alerts: a retail store, a warehouse, a factory floor. Those alerts land on a phone. The person holding it is usually not a security analyst in a control room. It’s a store manager or a site head with a dozen other jobs.
At 3:14 AM their phone says Restricted Entry, Camera 7, Zone B. At 5:16 PM the next day it says Unknown Car, Camera 2, Zone F. It’s their own car. After a week of that, they stop opening alerts. And the one that matters arrives looking exactly like the dozens that didn’t.
90 to 99%
of alarms reaching monitoring centres are false or not actionable
About half
of target events caught by trained operators in a 90 minute video review study
200+
businesses running JARVIS, each with its own cameras and its own idea of normal
Industry figures, not Staqu’s: The Monitoring Association and CSAA data as cited by Actuate (2026), and a study of 42 video operators in Applied Ergonomics as cited by Security Journal Americas (2026). Client count from staqu.com.
WHAT I FOUND
Three reasons good alerts get ignored
01
Generic models, specific sites
A model trained on generic footage meets a store under neon at 8 PM and a warehouse at 4 AM. The problem isn’t that thresholds exist. It’s that tuning them for every site is ongoing work nobody on site has time for.
02
The invisible correction
When someone dismisses a false alert, does anything tell them it counted? If not, the feedback loop can work perfectly and still be invisible. Trust never builds, even while accuracy does.
03
Every alert looks the same
A forklift in the wrong lane and a stranger at the back door arrive with the same weight. When everything is urgent, nothing is.
USER GOAL
Only alert me when something actually happens.
BUSINESS GOAL
Stop customers churning out of frustration, keep the ones already wavering, and build a system that gets better on its own, so support load drops instead of rising.
THE OPTIONS
I had three answers. I built the one that lives inside the alert.
01
Confidence tiers and a visible feedback loop
Alerts arrive in tiers, so the system only interrupts for what it’s sure about. Every alert carries a way to say “this was wrong”, and every correction gets an answer.
BUILT
02
Zone and time rules
Let people draw zones on a camera and set rules per zone, because a loading dock at 2 AM and 2 PM are different places. Powerful, but it hands the tuning job back to the exact person who has no time for it. That’s problem 01 again.
LATER
03
A weekly noise report
Alerts this week against last week, false alert rate by camera, the noisiest cameras, the zones that improved. Great for the person who renews the contract. Useless to the person holding the phone at 3 AM.
LATER
Option 01 won because it works inside the moment people already live in: the alert itself. No setup screen, no report to remember to open. It’s also the most practical of the three: it ships on top of the alert feed that already exists, with no new setup screen and no new product to sell. And it quietly feeds the other two. More on that in the screens.
THE FLOW
Three taps from annoyance to proof
Tap “Tell us if not” under any alert. Pick the closest reason, or write your own. Submit. JARVIS answers straight away with what your correction changed. The whole loop happens on top of the alert feed, so nobody loses their place.

THE SCREENS
Four screens, one promise: your correction counts
01 · TRIAGE
Not every alert gets the same voice
Alerts are tiered High, Medium and Low, both as filters along the top and as a coloured tag on every card: warm for Medium, cool for Low. The camera still sits on the card, so a glance is often enough to judge. And the way to say “this was wrong” isn’t hidden in a menu. “Was this alert accurate? Tell us if not!” sits under every single alert.

02 · FLAG
What made this a false alert?
Five reasons, and each one points the model at a different fix. Wrong person identified is a recognition error. Normal activity for this time is a time rule. Wrong zone triggered is a zone rule. Known staff or visitor is an allow list. This is where option 02 sneaks back in: every tap collects the zone and time rules nobody would ever sit down and configure.

03 · EXPLAIN
Something else? Say it in plain words
Reasons can’t cover everything. “The unidentified vehicle was my own car” is exactly the kind of context a camera can’t see and a person knows instantly. Free text comes second, never first, so the quick path stays quick.

04 · PROOF
Noted. And here’s what it changed.
This is the screen the whole concept exists for. “JARVIS has noted this. Based on feedback like yours this week, 27 similar alerts won’t interrupt you again.” Not “thanks for your feedback” into the void. A number, tied to your correction, that you can hold the system to. And “feedback like yours” makes it a team effort: everyone on the site teaches the same JARVIS.

TRY IT
Flag a false alert yourself
Tap “Tell us if not” under an alert to start.
WHAT I’D CHANGE
Built in a day, so here’s what I’d fix on day two
PROBLEM
The tiers lost their verbs.
FIX
My first sketch called them Act now, Review soon and FYI. The screens shipped as High, Medium and Low, which only say how loud an alert is, not what to do about it. I’d put the verb back on the tag.
PROBLEM
Checkboxes under a “pick the closest reason” question.
FIX
The control says “pick many”, the copy says “pick one”. I’d switch to single select rows, since one clear reason is more useful to the model than three vague ones.
PROBLEM
“27 won’t interrupt you again” is also a risk.
FIX
A real intruder can look a lot like your own car. Suppressed alerts shouldn’t vanish. They should drop into a quiet FYI queue you can check, with an Undo on the confirmation screen.
PROBLEM
It’s mobile only.
FIX
The person who renews the contract is often a security head watching many sites from a desktop. That’s where the weekly noise report (option 03) belongs.
TRADE OFFS
Every decision closed a door
01
Visible proof over silent learning
Showing the result of a correction is what builds trust.
WHAT IT COST
once you promise a number, the model has to be able to back it.
02
Feedback on the card, not in settings
The correction lives where the frustration happens.
WHAT IT COST
one more line on every single alert.
03
Reasons first, free text second
Fast for the common cases, and structured enough for the model to use.
WHAT IT COST
the list only covers the failures I could predict.
04
Tiers instead of mute
Nothing gets switched off, low confidence alerts just get quieter.
WHAT IT COST
Low alerts still arrive, so the feed never goes fully silent.
05
The alert moment over the setup screen
I designed for the person at 3 AM, not the admin configuring cameras.
WHAT IT COST
zone and time rules only build up slowly, one flag at a time.
WHAT CHANGED
The same Unknown Car, two different tomorrows
Before (dismiss and hope)
Unknown Car, Camera 2. It’s my car. I swipe it away. Nothing tells me that mattered. Tomorrow at 5:16 PM, same car, same alert, and I trust JARVIS a little less.
After (flag and see)
Tell us if not. Something else: “The unidentified vehicle was my own car.” Submit. JARVIS answers: 27 similar alerts won’t interrupt you again. Tomorrow at 5:16 PM, my phone stays quiet, and I know why.
The number I’d measure first is how often the same false alert comes back after someone has flagged it. Today that number is invisible, so nobody can tell whether flagging helps. If this design works, repeat alerts per camera fall every week.
FOR THE PERSON ON SITE
Share of alerts actually opened, climbing back up because what’s left is worth opening. Share of false alerts that get flagged, which tells me the loop is easy enough to use at 3 AM.
FOR THE BUSINESS
Support tickets about false alerts, falling. Renewals from accounts that flag often against accounts that never do, because the first group can see JARVIS getting better and the second can’t.
Repeat false alerts after a flag
THE ONE METRIC THAT MATTERS
WHAT I LEARNED
In B2B AI, trust is a UI problem before it’s a model problem
I came in thinking the job was to reduce false alerts. But the model can get better every week and the customer can still churn, because nobody on site ever sees it get better. The most valuable screen in this project isn’t a dashboard or a filter. It’s one sentence after you submit a correction.
Having one day also taught me something. I found the single screen that carries the idea, the Noted screen, and built everything else backwards from it. The tiers, the reasons, the free text all exist to earn that one sentence.
More work
Check out more work
ixigo Trains · The four hours nobody designs for
Twelve · Plans in years, not months
Google Pay Split and Request



