DUAA KHALID / AI PRODUCT + GTM
NEW YORK · OPEN TO FORWARD-DEPLOYED AI ROLES

I get software into companies that are not ready for it.

Five years in luxury beauty, running accounts across MENA. I built the tools my market needed and the field team opened them every day. Nobody made them. All of that in an industry that is still working out what AI is for.

MEng, Data Science & Decision Analytics. Cornell Tech, expected May 2027.

LVMH Dolce & Gabbana Cornell Tech 2021 to now
Duaa Khalid in front of a Dolce and Gabbana sign
DOLCE & GABBANA / 2026

ABOUT / DUAA KHALID

I learned product work on store floors, not in a lab.

Fifty five beauty advisors reported through me, and their salaries and commissions moved with their own numbers. When somebody on that floor had a bad month you saw it in the account inside a week. Everything I have built since assumes nobody will be made to open it.

Five years in commercial and GTM at LVMH, then Dolce & Gabbana Beauty across MENA. I still like the industry. What it could not give me was the technical side of the work I had already started doing on my own, so I am now doing an MEng in Data Science and Decision Analytics at Cornell Tech, finishing May 2027.

WHENEVER SOMETHING GOES WELL “Okay, we can still double this.”

HOW I WORK

I need to know what everyone else's job actually is.

Not out of curiosity. If I do not know what lands on somebody else's desk and what they are waiting on, I cannot tell why my own work has stopped moving. So I go and ask, and the answer is rarely what the situation looked like from the outside. It is slow, and on the tracker it is most of the reason the thing got used.

  1. WHAT IT LOOKED LIKE

    Somebody in the chain is being difficult.

    WHAT IT ACTUALLY WAS

    A process I could not see from where I was sitting.

    The platform we already licensedChanging a monthly target meant emailing the vendor's IT team and waiting for somebody there to do it. Nobody in that chain was slow. The permission to make the change sat somewhere I had not thought to look.

  2. WHAT IT LOOKED LIKE

    A pilot is there to prove the thing works.

    WHAT IT ACTUALLY WAS

    It is there to collect what they hate while changing it is still cheap.

    Tracker pilot, one weekThree advisors and their supervisor, on real data in a real month, before anyone else saw the app. The supervisor's commissions calculated themselves instead of by hand, and that week is where the fix list came from.

  3. WHAT IT LOOKED LIKE

    Nobody is using it, so I built it wrong.

    WHAT IT ACTUALLY WAS

    They could not reach it. Nothing in the app was broken.

    Tracker launch, AprilRegistering needed an email address, and not everyone has one they can get into. I sat with the advisors who were stuck instead of sending another set of instructions, then ran WhatsApp alongside the app for a month while people came online on their own dates.

CAREER PATH

Every step started as somebody else's broken system.

NOW / FROM 2026

HOMEBASE LABS

Co-founder, product and GTM. AI voice and WhatsApp agents that dispatch property maintenance, built around the UAE rental compliance rules the agent has to respect. The case study is the part I would defend: the call context is a deterministic template, not a model.

2026–2027

CORNELL TECH

MEng in Data Science and Decision Analytics. I have been building these tools alongside the day job, on free tooling and whatever I could teach myself in the evenings. I want to learn to do it properly.

2024–2026

DOLCE & GABBANA BEAUTY

Key account management across thirteen markets in travel retail MEA, with 55 beauty advisors reporting through me. Nobody could tell an account was slipping until it already had, so I built the account health layer that benchmarks the accounts against each other and flags the ones at risk. It is now the primary decision input across all thirteen. One of those markets was my own account, and it grew roughly 6x in eighteen months, which is where the sales tracker came from.

2021–2024

LVMH · MENA

I started out tracking out of stocks and aged inventory by hand across the MENA accounts. Then I became internal product owner for the sell-in tool the account managers used every day. I wrote the requirements and ran the roadmap, then sat in feedback loops with those account managers for two years. It was the first time I was the single point of contact across retailers, marketing, demand planning and supply chain.

6x SALES GROWTH ON MY ACCOUNT · 18 MONTHS
55 BEAUTY ADVISORS ON MY TEAM
02 DAYS TO A RUNNING BUILD

SELECTED WORK / 2024–2026

Three systems. Three different kinds of broken input.

One at a time. Use the arrows or the numbers to move through all three.

01 / FIELD DATA SYSTEM

DOLCE & GABBANA BEAUTY · TRAVEL RETAIL MEA

Beauty Advisor Sales Tracker

We already licensed a platform for this. It could not be configured by the person accountable for the numbers, it was priced per seat, and spending more on it would not have fixed either problem. So I built the replacement. It runs on about two dollars forty per user a month at paid tiers, and at our volume it costs nothing, so there was no spend to approve and no case to make. Two days to a running site on Supabase, added to the advisors' home screens from the browser rather than shipped through an app store, and piloted with three advisors and their supervisor for a week before anyone else saw it.

PWA · PYTHON INGEST · DATA VALIDATION · FIELD ROLLOUT

VIEW DASHBOARD EVIDENCE →

I had backfilled everything up to launch, then advisors came online on different dates. For a while we were both entering the same advisor's days, with no way for her to see mine. The dedup key decides whether two rows are one event or two.

ba_name entry_date store shift sales_amount status
Hana M. 2026-03-14 Cairo Morning 1,840 kept
Hana M. 2026-03-14 Cairo Morning 2,310 removed
ROWS2 SALES COUNTED4,150 DAYS WORKED2

The data columns and the four-column key are the real schema. Status is added here to show what the pipeline did, and the rule shown keeps the first occurrence. Values and advisor name are illustrative. The same export arrived with 38, 38, 40 and 37 columns across four consecutive months.

Try it · pick a city, type a target, and the card recalculates. Demo data, real arithmetic.

Performance Review Targets & Commissions Daily BA Sales BA Profile Shops Missing Sales Animations
MONTH
Jul 2026
CITY
All cities
MONTHLY TEAM TARGETS Jul 2026 · 30 days in month
July 2026
Per BA target

Demo data, and it is live. Set a target and the card recalculates.

Cairo ✓ On Track
$38.9k of $39k target · Jul 2026
Remaining$0.5k
Daily needed–
Last 7-day avg qty/day6.4 pcs
Hurgadah ✓ On Track
$42.0k of $40k target · Jul 2026
Remaining$0.0k
Daily needed–
Last 7-day avg qty/day8.3 pcs
Sharm ↓ Behind Target
$26.9k of $31k target · Jul 2026
Remaining$3.9k
Daily needed–
Last 7-day avg qty/day5.7 pcs

02 / AI PRODUCT

HOMEBASE · CO-FOUNDED, 2026

Homebase

The obvious answer is to put a model on the call and let it write the script. But the variation was never the useful part. The inconsistency was, and what a vendor needs on the other end of the phone is the same eight fields in the same order every time. That is a template problem, not a model problem, which is why the default path is a plain string builder and the model only runs behind a flag. Six Firestore collections sit behind it, and it ran a pilot with a UAE real estate company.

VOICE AGENT · DETERMINISTIC TEMPLATE · FIRESTORE

VIEW CONCEPT PAGE →

Change the ticket. The block keeps its shape, because the shape is the product.

build_context() · deterministic path
The Homebase concept page showing the assistant handling a tenant maintenance request

03 / PERSONAL BUILD

MOVE-OUT SALE · DUBAI, 2026

Move-Out Sale Site & Tracker

A personal build, and a small one. A catalog fixes the messaging and does nothing about the part that was costing me money. I was pricing against store catalog pages rather than against what I had paid, and those pages were wrong by more than double. So one Python script now generates the buyer facing catalog and the profit tracker from a single item list, and prices against what was paid, which means a saving cannot be displayed unless it is real.

PYTHON · EXCEL MODELLING · STATIC SITE · PRICING RULES

VIEW LIVE CATALOG →
The live move-out sale catalog showing furniture listings

FIELD NOTES

They do not work for the brand. They work for you.

Adoption is won on trust, long before anyone sees a feature.

I had 55 beauty advisors under me and I was responsible for their salaries, their commissions, and whether they were motivated enough to get through a shift. That last one is not a soft thing. If somebody on that floor is having a bad month their sales change, and that changes my numbers. You feel it inside a week.

People join because they love the brand. They do not go above and beyond because of it. They work for the person in front of them, and if they trust that person they will try something new. If they do not, they will do exactly what the job description asks and nothing more. Nobody stays late for a logo.

Which is why I think adoption is decided before the product is opened, not inside it.

The tracker was a lightweight sell-out dashboard that replaced several spreadsheets and gave the advisors one place to see their daily targets, their sales and their commissions. When I put it in front of eighteen people, nobody had told them to use it. No mandate, no deadline, no manager checking. They used it because I had spent two years building the kind of relationship where if I said this will make your life easier, they believed me enough to try it once. Trust got them to open it. The product earned the second visit.

That is the part I think transfers. Forward deployed work means shipping to people who do not report to you, inside a company that is not yours, who already have a full job. You cannot mandate them. You have to be the person whose recommendation is worth trying. Every decision on that rollout came out of it: the home screen instead of the app store, WhatsApp kept running alongside, and refusing to set a cutover date until people were ready. None of that is clever engineering. It is what it costs to not burn the trust you already have.

It is worth saying what that means for anyone selling into a company like the one I worked in. Your competition is not always another vendor. It is somebody inside the account who can put a working version together in an afternoon, because the tooling got that cheap, and who understands the workflow better than you do. I was that person. If the problem is small enough and close enough to the business, your customer will build it themselves.

The lesson was never that I built a better tracker. It was that trust is part of the product.

QUESTIONS & CAVEATS

Your questions answered directly.

The questions that come up when someone reads this site, including what the AI work in these projects actually is.

01. Why did you build this instead of using the tool you already had? +

I owned a travel retail account in the Middle East and it was growing, roughly 6x over about 18 months. The advisors were sending me their daily sales over WhatsApp and I was retyping them into Excel by hand.

We already licensed a field-sales platform. It worked for the large markets and not for mine. Loading a monthly target meant emailing the vendor's IT team and waiting for someone there to do it, the dashboard did not show what I needed, and the per-seat cost could not be justified for a market that size. What replaced it runs at about two dollars forty per user a month on paid tiers, and nothing at all at the volume we actually needed.

Building the replacement cost nothing but my own time, so there was no spend to approve and no case to make. Two days from the decision to a running site, with Supabase behind it, which I learned on this project.

It was a judgement call, not a brief. Nobody assigned it and nobody was waiting for it. I decided, as the person accountable for the numbers, that the licensed tool was the wrong fit here. Worth saying plainly, because being handed a project is a different thing.

02. What went wrong in the rollout? +

The signup screen, and I did not see it coming.

The build was not the hard part. The advisors work off their phones, so rather than ship a native app I had them add the site to their home screen from the browser. I piloted with three advisors and the sales supervisor for a week, took their feedback, fixed it, then launched to everyone.

Then people could not get in. Registering needed an email address, and not everyone had an email account they could actually get into, or had ever signed up for anything online.

I did not force the change. WhatsApp stayed on alongside the app for a month so anyone still stuck could send numbers the way they always had, and I reconciled the two channels by hand. People moved across as they got set up, not because a cutoff date made them. Nothing was broken. I had just never thought to check whether the people I built it for could reach the login screen.

03. How much of this is actually AI, and how much did AI write? +

Less AI in the products than the word implies, and more AI in the building than most people admit.

The tracker and the move-out generator contain no models. Homebase has an opt-in model path behind a flag (USE_LLM_SCOPE=true) and a deterministic template builder as its default, because a string template was sufficient.

I built with AI, in Cursor and Claude Code, and I would not have shipped any of it otherwise. What I will not claim is that AI did the thinking. It will not tell you what makes two rows the same event, what a supervisor needs on a dashboard, or what a field team on phones can and cannot do. You have to know that first, and then the tooling gets fast.

04. Eighteen advisors is a small number. +

It is, and I am not going to dress it up. What matters is that nobody was required to use it.

The tracker was never mandated. Eighteen advisors across Cairo, Sharm and Hurghada opened it daily anyway. A mandated tool measures compliance. An optional one that people open every day measures whether it was worth their time.

For scope rather than adoption: I was responsible for 55 advisors at Dolce & Gabbana and for account health across thirteen markets.

05. What have you not done? +

I have not shipped a machine learning model to production. I have not worked at engineering scale: the largest user base here is eighteen people and the largest dataset is a few thousand rows. I have not managed engineers. Homebase's compliance layer is specified in its README and listed as out of scope in that same README, so it is a spec and I have not built it.

What I have done is take a broken operational input, decide what a valid record is, and get a field team in another country to work the new way when nobody had asked them to. That is the part I would defend.

On my own sites, nothing is broken right now. The move-out catalog could previously show "0 items available" instead of the real count. I listed it here while it was open and it is now fixed, reading 11. I leave resolved entries up rather than deleting them, because a list that only ever shows open problems tells you less than one that shows what happened to them.

What these numbers do not mean

  • The dashboard on this page is demo data: recreated in HTML with invented figures and invented names. No real advisor performance is published anywhere on this site.
  • January to March was backfilled: the app launched in April. The first three months of the 1,324-row window were loaded by me in SQL from existing records, not logged daily by advisors.
  • Removed, not rejected: the 29 were valid rows that collapsed against the dedup key. 1,324 in, 0 rejected. Nothing failed validation.
  • No cycle time measured: month-end turnaround improved and I never instrumented it, so I am not putting a figure on it.
  • 18 live, about 20 onboarded: two real numbers counting different things, not a rounding of one.
  • The honest move-out figure is 20 of 31: the sheet says 21 of 32 because a COUNTIF double-counts a synthetic bundle row.

CAPABILITIES

BUILD

  • Python
  • SQL, including backfilling three months by hand
  • Supabase, learned on the tracker
  • Firestore
  • Sites people add to a phone home screen

OPERATE

  • Field rollout and training
  • Key account management
  • GTM across MENA
  • Pilot first, then launch
  • Getting a team off WhatsApp

DEFINE THE RULES

  • Deciding what counts as a valid row
  • Dedup keys and validation rules
  • Commission and target logic
  • Account health benchmarking

WHAT'S NEXT

Building a forward-deployed or AI product team? Let's talk.

DK947@CORNELL.EDU LINKEDIN →
DUAA KHALID / 2026 NEW YORK