Open to connect

hi there, I'm

Anna Kostiuchenko

Data Analyst

I hold a Master's in Digital Business Analytics and work with DWH data, AI automation and business analytics. I use SQL, Python and Power BI to turn complex data into clear, actionable insights, and I like dashboards that tell a story at a glance.

Enschede, NL LinkedIn

8.5 GPAMSc Digital Business Analytics, cum laude

Business AnalystCustomer Growth Tribe at PUMB, current role

Marketing + Dataexperience in digital marketing and data analytics

Skills

The tools I work with every day.

SQLSQLQuerying & DWH data

BIPower BIDashboards & DAX

TbTableauVisual storytelling

XlExcelPivots & slicers

PyPythonAnalysis & automation

RRStatistics

My CV

Work experience, education and languages in one page.

PDF · 1 page

Anna Kostiuchenko, Data Analyst

My full CV with work experience at PUMB, J58 and Nedap, my degrees from the University of Twente, skills and languages.

Download CV PDF

Dashboards

Projects grouped by the tool they're built in.

Excel

Pivot tables, slicers and interactive charts

Sales analytics

Sales Dashboard · Germany 2021–2023

An interactive sales overview for a food wholesaler, built on about 19,000 order lines. Slicers for year, region and city filter every chart at once, so a manager can go from the national picture to a single city in two clicks.

1.26MBerlin revenue, the top city

35%East, largest regional share

3 yrsmonthly revenue trend

What it shows

Revenue peaks every February–March, with 2023 reaching its highest month at 332K.
Coffee and Gouda are the best-selling products; Beverages lead all categories.
A bubble map of Germany shows revenue by city, from Bremen (228K) to Berlin (1.26M).

Download workbook .xlsx · 1.4 MB

ExcelPivot tablesSlicersMap chart

Tableau

Calculated fields, statistical testing and storytelling

A/B testing

Campaign Performance · A/B Test 2025

A marketing experiment report comparing Control and Test variants of two campaigns across 323.5K audience entries. It tracks CTR month by month, compares conversion rates, and checks whether the difference is real with a two-proportion Z-test.

+15.3%conversion uplift, Test vs Control

9.10%best variant: Campaign B Test

p=0.0064below the 0.05 threshold

Key insights

Test variants beat Control every month on CTR, rising from 3.75% to 4.8% by December.
Campaign B's Test group converts best (9.10% vs 8.35% for Campaign A).
Recommendation: prioritise the Test variants, especially B, and confirm with a live randomized experiment before full rollout.

View interactive dashboard Download workbook .twbx

TableauCalculated fieldsTwo-proportion Z-testKPI tiles

Public data · SQL + Tableau

Dutch Housing Market · 1995–2025

How much have Dutch house prices risen, and where? I joined and cleaned 30 years of official CBS data for all 342 municipalities in SQL, then built a Tableau dashboard where you can compare any municipality with the national average.

~5×national average price, 1995 to 2025 (€480K)

Flevolandfastest-rising province since 2015

−22%Enschede below the national average

Key insights

Prices dipped after 2008 and again in 2022–23, but the long-term rise is steep.
The fastest risers are commuter towns around Amsterdam and in Flevoland, such as Almere, Purmerend and Lelystad.
Growth uses 3-year averages (2013–15 vs 2023–25), so one unusual year can't distort the ranking. Diemen still tops the list, partly because of a jump in 2018 that likely reflects more new-build sales.

Data: CBS table 83625NED, average sale price of existing homes. This is not a price index, so changes in which homes are sold affect it.

View interactive dashboard Download workbook .twbx

TableauDuckDB SQLLOD expressionsParametersMaps

Real public data · AI + Tableau

Driver Voice · Logistics Warehouse

What do truck drivers say about loading and unloading at a large logistics warehouse in the Netherlands? I collected 162 public Google reviews in 15 languages and had Gemini classify each one by reviewer type, topic and sentiment, and pull out the reported waiting time. A daily Apps Script trigger tags new reviews automatically, and Tableau turns the result into a dashboard.

74%of driver reviews are positive

80%out within an hour (of those who mention a time)

30 minmedian reported wait

Key insights

Fast loading and friendly staff are what drivers mention most, and mostly positively.
Check-in and paperwork has the highest share of negative mentions, together with an unanswered intercom at the gate.
The next step would be to check reported waits against dock-appointment and gate data from the warehouse systems.

Data: public Google Maps reviews, collected October 2026, reviewer names removed. Dates are approximate, and the wait-time figures rest on the 15 driver reviews that mention a time. Independent analysis, not affiliated with the company.

View interactive dashboard

TableauGemini APIApps ScriptGoogle SheetsText classification

Retention · SQL + Tableau

Customer Retention · Online Retail 2009–2011

Who comes back after their first purchase, and who brings the revenue? I cleaned about 1 million order lines of a UK online retailer in SQL, built monthly cohorts and RFM segments, and visualised them in Tableau.

~1 in 5new customers buy again the next month

9%Dec 2010 newcomers back next month, vs 21% avg

70%of revenue from 27% of customers

Key insights

December shoppers are the least loyal, while lapsed customers from every cohort return each November, so retention is seasonal.
Champions (27% of customers) bring 70% of revenue; Lost customers (28%) bring just 4%.
Recommendation: a January win-back campaign for holiday first-time buyers, and targeting the At-risk segment (12% of customers, 9% of revenue) before they are lost.

Data ends 9 Dec 2011, so each cohort's last month is incomplete. The SQL is in the SQL section.

View interactive dashboard Download workbook .twbx

TableauDuckDB SQLCohort analysisRFM segmentation

Power BI

DAX, data modelling and performance monitoring

Campaign funnel · demo data

Promo Campaign Funnel · Birthday & Holidays

A report inspired by my real work at PUMB. To keep security in mind, it uses fake data. It follows two promo campaigns from audience to offer. A matrix breaks every campaign down by banner and shows each funnel step, from moving to the Credits section to seeing a banner, clicking it and using the offer. Slicers filter by month, campaign, banner, test group and product.

18,600clients across both campaigns

32%moved on to the Credits section

411clients used the offer

What it shows

The funnel matrix uses in-cell data bars, so the strongest banners stand out at a glance.
Conversion is compared between the control and main groups for each banner.
A Z-test per product checks whether the main group really converts better. None of the differences were significant (total p = 0.14).

Power BIDAX measuresMatrix & data barsZ-test

SQL

The queries behind the dashboards, written in DuckDB

Customer retention

From 1M raw order lines to cohorts and RFM segments

Data-quality checks, cleaning, monthly cohorts and customer segments, ready for Tableau.

CTEsWindow functionsNTILEFILTER

View on GitHub

Dutch housing

30 years of CBS data for 342 municipalities

Joins CBS code tables, handles merged municipalities and ranks growth with 3-year averages.

JoinsRANK()UNION ALLSanity checks

View on GitHub

AI & Agents

AI that does real work: answering data questions and triaging feedback

Data agent · banking

A DWH agent that sizes target audiences on demand

Built at PUMB. The approach is shown here; the bank's data stays private.

ProblemManagers wanted to test target-audience ideas, like how many clients match certain criteria, but had no SQL skills or DWH knowledge. Every question went through a small analyst team, so testing hypotheses was slow.

ApproachI duplicated the relevant database, anonymised and aggregated it, and built one wide table of client attributes. Then I connected that table to an AI agent that turns a plain-language question into a query.

ResultManagers can ask in their own words and get the size of a target audience in moments, so they go into a campaign with clear expectations, and analysts spend less time on one-off counts.

DWHSQLData anonymisationWide-table designLLM agent

AI automation · live demo

Fietsly: AI feedback triage you can try

A working pipeline for a made-up bike rental company. Feedback from a form is tagged by Gemini and appears on a live dashboard.

ProblemSupport teams drown in free-text feedback. Reading every message to spot safety issues or wrong charges is slow, and urgent cases get buried.

ApproachGoogle Form → Sheet → Apps Script sends each response to Gemini, which returns sentiment, topic, urgency and a one-line summary. Only allowed values are accepted, failed rows retry automatically, and a privacy-safe copy without emails feeds the dashboard.

ResultEvery response is triaged within seconds, urgent issues surface in an action list, and the model resisted a prompt-injection attempt hidden in the feedback.

Open live dashboard Try it: leave feedback

Google Apps ScriptGemini APIPrompt designAutomationChart.js

Research

How AI and LLM agents shape decisions in firms

PAPER
2026

Synthetic Firms: Employing LLMs as Organisational Simulacra

Johannes Dahlke, Flavio Calvino (OECD), Anna Kostiuchenko · SSRN working paper, May 2026 · 40 pages

Can large language models, guided by economic theory and calibrated with real data, imitate how firms innovate? We built natural-language profiles of German firms and compared an uncalibrated model with two versions calibrated using web-derived indicators and patent data. The uncalibrated model mostly overpredicts innovation; both calibrated versions are much more accurate and match real patterns across industries, including in a validation on a later period.

Read on SSRNDOI 10.2139/ssrn.6681238

LLMsSimulationSynthetic dataInnovationCalibration

MSc
THESIS

How can theory-grounded LLM-based agents synthetically represent firms to predict real-world, firm-level technology adoption?

MSc Business Administration, Digital Business Analytics · University of Twente, July 2026 · supervised by Dr. J. Dahlke and Dr. G. Sedrakyan

I built LLM agents (GPT-4o mini) that each represent a real firm and predict whether it adopts AI tools. I tested them on 3,531 German automotive manufacturers, compared six management theories as the agents' reasoning lens, and grounded them with a peer signal from the 25 most similar firms. Results were validated against real adoption data and official ZEW, ifo and OECD statistics.

Without grounding, the LLM barely beats chance and leans almost entirely on firm size. Pairing theory with a peer-adoption signal raises accuracy substantially.
Diffusion of Innovations and Porter's Five Forces were the strongest lenses. Extra firm data helped only when the theory could use it, and splitting complex predictions into simpler steps worked better than one pass.

Read the thesis

PythonOpenAI APIKNNDesign-Based ResearchAUC evaluation

BSc
THESIS

How can Generative AI influence decision-making in early stage startups operating under high uncertainty?

BSc International Business Administration · University of Twente, June 2025

A mixed-methods study (a survey plus interviews with early-stage founders) on how GenAI tools shape startup decision-making. Most founders use a hybrid approach, relying on GenAI for marketing, technical work, ideation and research. Despite worries about hallucinations, trust and prompting skills, they see it as a practical way to work faster, cut costs and experiment more. The thesis proposes a framework for intentional GenAI adoption built on trust calibration, prompt literacy and scalable integration.

Read the thesis

Mixed methodsSurveyInterviewsGenAI adoptionEntrepreneurship

Let's talk data

I'm happy to chat about analytics, dashboards or AI automation.

linkedin.com/in/anna-kostiuchenko11