Nate DouglasSenior Product Designer, home

Friday is booked.
Monday is only proposed.

A design challenge for Krane. A working internal tool that turns supplier messages, site updates and approvals into one clear next decision.

Role
Solo designer and researcher. I wrote the code myself, using AI coding tools
Timeline
September 2026
Context
Design challenge for Krane
Study
Five construction professionals, six tasks each
23 of 30successful task attemptsOne unmoderated round: 5 construction pros, 6 tasks each
5.1 / 7average ease ratingSame round, rated by participants
5 of 5told the recorded booking apart from the conditional proposalSame round, read from written answers
How these were measured

One unmoderated round with five construction professionals, recruited through UX QuestionPro and paid $70 each. I read their submitted responses rather than watching sessions. No retest yet, so there’s no before and after.

Krane delivery detail for CHLR-02: Friday September 18 booked, Monday September 21 proposed at $450 more, with four requirements outstanding
What people said

From the team behind Krane.

Eshan Jayamanne
“Nate did a design challenge for us at Krane, and he didn’t stop at Figma. He built a working app in React, and turned it around fast. Our whole product lives or dies on whether people trust what the AI tells them, and he nailed that right away. Would recommend him to anyone building AI products.”
Eshan Jayamanne, Founder and CEO of Krane
The brief

Every message changes something, and it’s hard to tell what.

Krane asked me to use AI to make sense of construction information while keeping people in charge. A summary that sounds confident can hide the one condition that matters.

  1. A production date can be mistaken for an arrival date.
  2. A receiving window can be mistaken for a transport booking.
  3. A lower price can look better before freight is added.
AHU-04 procurement conditions: Revision B approved, Revision C still under review, with the fabrication condition and cutoff beside the completion target
Rev B is approved and Rev C is still under review, so fabrication stays conditional and no delivery date is set. What’s still unclear is listed with its next check.
What I decided
Make checking easier, without hiding what’s still uncertain.

A review step only where a wrong reading could cost money, time or material. No checkboxes everywhere.

In testing, all five people could tell what was booked from what was only proposed.

Usability study

Five construction professionals, six tasks each.

Time on task by participant: five lines across the six tasks, in seconds
Time by participant, read alongside completion and decision quality. These tasks don't establish time savings against a baseline.
Participant breakdown: role, experience, each task outcome and session length for all five participants
Five profiles and 30 task attempts: 23 complete, 5 partial and 2 not complete.
Finding F3: reading a reply does not make the review step obvious, 2 of 5 participants
Freight review: the two-record review already exists. Next, check whether the acknowledgement and Apply action explain what changes.
Finding F4: receiving availability can be mistaken for readiness, 1 of 5 participants
Field handoff: receiving availability is supported. Access, unloading, written charge approval and carrier confirmation stay separate checks.
Internal tools

One queue for the people who run the job.

Coordinators track dozens of materials across email, a spreadsheet and phone notes. The tool is built around three jobs: find what changed, decide the next step, hand it to the right person.

What’s inside

  1. Work queueOne owner, one next action and a due time per material.
  2. Delivery reviewThe AI’s reading next to the original message, before anything changes.
  3. Supplier comparisonQuotes on the same basis. Freight applies only after review.
  4. AI quality reviewCorrections tested against current behavior before they go live.

All four share one demo library of status, evidence and review patterns.

Needs a decisionNorthline and Harbor · 4 items

Sorted by what blocks a date

Work queue: four materials, each with what changed, the next action, owner, due time and status
MaterialWhat changedNext actionOwnerDueStatus
AHU-04Air handling unitRev C still under review. Fabrication stays conditional.Clarify drawing and fabrication conditionsJamie DavisTue 12 PMBlocks a date
CHLR-02ChillerCarrier reply sounds positive. Four requirements still open.Confirm access and unloadingAlex MorganWed 3 PMNeeds checking
SWGR-01Switchgear, NorthlineEarlier $5,000 freight answer found.Review and apply freightPriya ShahThu 10 AMReady to review
SWGR-02Switchgear, HarborNo production slot for October 19.Ask Apex for separate availabilityPriya ShahFri 5 PMUnconfirmed
Every row has one owner, one next action and a due time. A row leaves the queue when someone reviews it, never just because a message arrived.

Who can do what

ActionProject coordinatorProcurement leadReceiving leadProject manager
Review an AI readingYesYesYesYes
Send a clarification requestYesYesYesYes
Apply a freight answer to a comparisonNoYesNoYes
Approve a charge, like Monday’s $450NoNoNoYes
Turn on a corrected AI ruleNoNoNoYes
Jamie coordinates the $450 approval but can’t give it. The screen names who can, instead of showing a button that fails.

Every change leaves a record

  1. Jamie Davis reviewed the carrier update. Friday stays booked.
  2. Jamie Davis sent a request to Alex Morgan with 3 sources attached.
  3. Priya Shah applied Beacon’s $5,000 freight from the September 10 email. Beacon now $99,000 delivered.
Who changed it, from which source, and what it replaced. When a supplier disputes a date, the answer is in the record.
Field delivery

A positive reply doesn’t mean it’s decided.

CHLR-02, a chiller for Northline Medical Pavilion. Friday 9 to 11 AM is booked. Monday is proposed at $450 more. Receiving space is supported, and four requirements are still open: route and access, unloading crew and equipment, written charge approval, and revised carrier confirmation.

Review delivery update: the carrier's original message beside the AI interpretation of booking, proposal and missing requirements
Review the updateThe original message sits next to the AI’s reading, so the reviewer compares the two. Saving the review never changes the booking.
Arrival options: Friday and Monday with aligned condition rows for weather, receiving, carrier and unloading
Arrival optionsThe conditions are lined up across the dates. With just a list of messages, people would have to do this comparison in their head.
Weather and receiving: possible rain from 6 AM to noon overlapping the proposed 9 to 11 AM receiving window
Weather and receivingPossible rain overlaps Monday’s window. That’s a reason to check, not a fact that Monday won’t work.
Who to check with: Alex Morgan for access and unloading, Jamie Davis for written charge approval, Maya Chen for carrier confirmation
Who to check withFour requirements, three people. Jamie coordinates the $450 approval but can’t give it.
Prepare clarification request: a draft to Alex Morgan carrying the relevant dates, conditions and sources
Prepare the requestThe draft pulls in the facts, so Jamie edits instead of rebuilding the history. Sending it doesn’t resolve anything.
Review Alex's reply: receiving space reaffirmed, access, unloading, written approval and carrier confirmation still open, with the original request below
Review Alex’s replyA partial reply on purpose: a message can sound positive and still settle almost nothing.
Updated material record: reply reviewed, Friday still booked, four requirements still outstanding
Material recordReply reviewed, Friday still booked, four requirements still open. A generic success message would have left Jamie guessing whether the plan had changed.

1 / 7Click a screen to see it full size.

Supplier comparison

Comparing two quotes on the same basis.

  1. Supplier comparison for SWGR-01 before freight review: Apex $100,000 delivered, Beacon $94,000 before freight, with a break-even view
    Situation
    For one switchgear package, Apex was $100,000 delivered and Beacon $94,000 before freight. Ranking those two numbers side by side would be comparing two different things.
    Call
    Keep the price basis visible and show the break-even: $6,000 of freight makes the totals equal. The only next action offered is reviewing Beacon’s earlier freight answer.
    Result
    Instead of a misleading ranking, the missing number turns into a specific question.
  2. Review earlier freight answer: the $5,000 freight answer and the later terms confirmation beside their interpretation and arithmetic, acknowledgement checked and Apply enabled
    Situation
    Beacon had already quoted $5,000 freight. A later confirmation referred back to it and added terms and an offer deadline. The answer was already there. It just wasn’t usable.
    Call
    Show both records beside their interpretation and the arithmetic. A checkbox confirms both apply, then Apply updates the comparison. I kept reviewing and applying separate on purpose.
    Result
    Beacon comes to $99,000 delivered, $1,000 below Apex on that basis. Applying it doesn’t approve a purchase. In testing, 3 of 5 completed this step, which makes it the clearest thing to fix next.
  3. Production availability: Apex slot SG-41 assigned to Northline, Beacon's capacity not confirmed, Harbor needing a separate option
    Situation
    Apex’s slot finishes production the week of October 5, and Northline needs the package on site October 12. Beacon’s capacity is unconfirmed, and Harbor needs its own option for October 19.
    Call
    Keep availability as its own check, with a connector that lands clearly on the project the slot supports. Ask for separate availability instead of quietly reallocating a commitment.
    Result
    A cheaper option never looks available, suitable or approved just because it’s cheaper.
When AI fails or is wrong

Planning for when the AI can’t finish, or gets it wrong.

AI review unavailable: the entered source and context kept, with options to continue manually or try again
When AI review can’t finish, the source and context stay and the person continues by hand. A failure never looks like it worked.
Compare rule results: the proposed correction against current behavior across the test cases
A correction becomes a proposed rule, tested against current behavior before anyone turns it on. Future updates only.

I ran a real test: 24 model calls scored by an AI judge against a saved rubric. Missed transport handoffs fell from 3 of 10 to none, and rubric failures from 8 of 60 to 4. Four citation errors remain.

What I dropped

Six things I left out, and why.

Checkboxes on every step

A review step sits only where acting on a wrong reading could cost money, time or material. Everywhere else there is no checkbox.

A price ranking

Two quotes on different bases don’t get ranked. The missing freight number becomes a question to answer first.

One-step review and apply

Reviewing never changes a booking or approves a charge. The cost: 3 of 5 finished this step in testing.

A generic success message

After a partial reply, a plain “done” leaves the coordinator guessing. The record says what’s still open.

AI fixes that go live at once

A correction becomes a proposed rule and gets compared against current behavior before anyone turns it on. It applies to future updates only.

A time-saved claim

The study had no baseline, so I don’t claim any time saved. That waits for a pilot against how a customer works today.

Not built yet: queue rows that open the message they came from, and filtering the queue by owner. Both are fast follows that don’t block.

Craft

Small fixes that changed what a screen meant.

Krane Demo Library index: colors, typography, layout, icons, buttons, inputs and selection, patterns and prototype guidance
A browsable demo library. The same status, evidence and review patterns carry across every screen, so a new internal tool starts from them.
The delivery screen after review: a green check mark reading Delivery update reviewed sits in the banner, and the header says Delivery update reviewed
Reviewed, and it stays reviewedThe mark survives opening sources, notifications and edit panels.
One slot, two project needs: a line runs from the Apex SG-41 slot to Northline Medical Pavilion, marked Recorded assignment; Harbor Research Center is marked Separate slot needed
The connector lands on the projectWhich project a production slot supports, at a glance.
Freight source evidence panel: the original records with the key phrases highlighted, next to the interpretation with buttons to view each source
Sources open the evidenceAnd close back to the same screen and scroll position.
Built in code

The prototype, running in code.

I built Krane as a working app in React and TypeScript, with its own AI runtime and automated tests. Here it reviews a carrier email, keeps Friday as the booking and lists the four things still missing.

The bar it has to pass to ship

“Ready to ship,” written as checks that read the demo library’s tokens and run on the work queue in this page.

Has to pass

  • Colors come from the libraryPass
  • Type uses the library’s sizes, weights and fontPass
  • Spacing stays on the 4 to 48 scalePass
  • Every status is an approved label with an iconPass
  • Every row has an owner, a next action and a due timePass
  • Text contrast is at least 4.5:1Pass
  • What the AI read never claims certaintyPass
  • No em dashes in the copyPass

Fast follow, doesn’t block

  • Each row opens the message it came fromOpen: no source links yet
  • The queue can be filtered by ownerOpen: not built yet

8 of 8 pass at desktop and 390px wide. The first run found 12 colors, the font and 4 spacing values off the library, so I moved the queue onto the shared values.

Looking back

Where people still got stuck.

What I’d change

Watch sessions instead of only reading submitted answers. Unmoderated responses told me who got stuck, but not always why.

What I’d do next

Fix the three places people stumbled: drawing requirements, the freight review-and-apply step, and the difference between receiving availability and unloading readiness. Then retest.

What I’d validate

Run a small shadow pilot next to how a customer’s team works today. Main measure: correct next actions. Guardrail: missed conditions. A 20% cut in whole-task time is a target to test, not a result.