Marina Havlis
Case study · Regional supply optimization

I took my analytical approach into a category I'd barely touched, and ran it end to end.

TL;DR

A chain of scripts that scans a region's listings, finds the ones under-converting or bleeding margin, and hands each fix to the account owner or the content team. I picked the KPIs, designed the flow, and had the cases in the reps' hands in two weeks, my second month acting as a category manager with no prior experience. Two of the five signed the week after, three weeks from start to first signature. This is the whole thing, end to end: the scan, the fixes that went out, what signed and what didn't, and where I'd take it next.

Marina Havlis · Data & Analytics Consultant · 2026

508
Listings scanned
5
Selected from 155 flagged
2 of 5
Signed (3 weeks start to signature)

How it went

The order I tend to work in.

It's the same six steps I'd use on any business problem, this time on a category I'd barely touched before. The tag on each step is the part of my skill set it leans on.

01
SignalAnalytical

A review flagged a margin metric sliding month after month. I wrote down a few reasons it might be doing that and tested each one. The one I expected (a handful of big players dragging down the whole pool) didn't hold up, so I ruled it out as the cause and built partner type into the scan along with other conditions, so it doesn't push the same fix on everyone.

02
PeopleCollaboration

Reps get dozens of automated cases thrown at them every day with no context, so they learn to ignore them. Before any of mine went out I sat down with the regional manager to make sure he personally asked three of his reps to prioritize those cases, and with the content team manager to agree the format of the page audits and make sure her team had the capacity to take them, so my cases didn't join the pile nobody reads.

03
MapAnalytical

I built the scan that scores every listing against its peers in the same market and category, so an underperformer shows up next to the listings it's actually comparable to, not against the whole region.

04
DiagnosisAnalytical + AI

The scan flags four groups. Two are in scope here, each down its own track: the ones making us too little per view, and those converting below their peers. The other two are out of scope, handled by separate tasks: a beyond-redemption group to cut, and one with a refund problem to chase down.

05
FixAI + Project management

The SQL is written by AI inside a Keboola transformation, an interface I know well, and I read it before it runs. I checked the methodology with a principal analyst who'd built things like this before. What comes out the other end is a full memo per partner, detailed enough that a rep can walk straight into the room with it.

06
MeasureAnalytical + Project management

The change audit waits 30 days for clean before-and-after data, then reports whether the margin fixes held. Anything sooner is just noise when the effect takes weeks to show up. The conversion track I didn't need to build; the content team already tracks their own pages at 30 days. The loop is in the design from day one, so the measured before-and-after is the next checkpoint.

The build

Four smaller scripts, each doing one thing.

I didn't want a single agent that "finds bad listings and tells me what to do." It turns into a black box I can't debug. Four smaller scripts, each doing one thing, are easier to test, swap out and explain.

1. Scan
Supply scan
Scores every listing against its peers in the same market and category. Sorts them into four groups: two go down the tracks here (low margin per view, below-peer conversion), two to separate tasks (cut the beyond-redemption listings, chase the refund problems).
↓   splits into two tracks   ↓
2. Margin track
Restructuring
A per-partner proposal, new terms and pricing, with the impact modeled for both sides. Goes to the account owner.
3. Conversion track
Listing-page audit
Reads the page and writes a per-listing fix list, positioning, options, how the price is shown. Goes to the content team.
↓   both feed forward   ↓
4. Change track
Change audit
Once a change is live, waits 30 days for clean before-and-after data, then reports what actually moved.

Each script ends in a real document someone can act on. The layout is exactly what shipped; the data is generated to prove the concept.

Supply scan diagnostic, generated-data sample
1 · Supply scan
Renegotiation memo, generated-data sample
2 · Renegotiation memo
Page-audit memo, generated-data sample
3 · Page-audit memo
Change audit scoreboard, generated-data sample
4 · Change audit

What it took before, and after.

The scripts took over the slow bit, pulling and formatting all those reports. The judgment calls stay with me.

Region scan
2–3 days
→ now
minutes
Listing audit
~45 min each
→ now
~5 min review
Restructuring
~3 hrs each
→ now
~15 min review

Roughly two working days a month, on an assumed cadence of four region scans, ~50 audits and ~30 restructurings a year. The hours that went into pulling reports go into the partner conversations instead.

Where I come in

I hand the rep the argument.

Most automations I encountered end by dropping a task on the rep: performance is low, go fix it. Not much use, no numbers, nothing to walk in with.

So what I hand over is a memo built for the negotiation: where the listing stands, what's pulling it down, the terms I'm proposing, and what they do for the partner and for us. The rep walks in with the case already made and does the persuading.

Renegotiation memo, generated-data sample
One memo per flagged partner, terms and both-sided economics in one place. Real format; data generated for the sample.

The result

What closed.

Two of the five signed within three weeks. The others said no, which was fine. The rejections were useful: they fed a follow-on analysis on which accounts are worth pushing harder, even at the risk of ending the partnership. The signals it read: customer reviews, volume trends, signs of undercutting us (a lower price direct than on the platform), which sent me back to tighten the scan. The AI-assisted kind of analytics is mostly iteration, I didn't expect to get it right in one pass. The account owner still walked into each renegotiation already knowing what changed for the partner and what changed for us. The report did that part.

What it shows

I'd never worked as a category manager before this. The work underneath turned out to be the same as any business problem: find where it leaks, work out the fix with the people who own it, build the thing, check it held.

Building the tools doesn't take long anymore. Where my time goes now is picking the right problem and checking the fix actually worked. On this one, two of the five deals signed in three weeks, in a role I'd never done.

If you're hiring for that kind of work, I'm open to short- and medium-term projects, Prague or remote. marinahavlis.pages.dev
marina.l.havlis@gmail.com