| Tool | Level | Proficiency |
|---|---|---|
| Power BI | Advanced | |
| Tableau | Advanced | |
| Google Looker Studio | Advanced | |
| Metabase | Intermediate | |
| Adobe Analytics | Intermediate | |
| Salesforce CRMA+ | Intermediate |
Four end-to-end projects from the neuefische AI Project Management programme, 2026
Each project ran on a real board with written success criteria agreed up front. Results are reported against those criteria, limitations included.
Problem: Lab analysts checked every weekly instrument run by hand against SOP thresholds, which was slow and led to inconsistent decisions. SpecTune answers “is the machine ready for today’s samples?” from the raw signal alone.
What I built: SpecXtract, a Python pipeline that turns ~300 GB of Thermo .raw files into per-run QC features. It pulls straight from the lab’s cloud share, runs extractions in parallel, and was validated 100% against the lab’s reference tool on 365,000+ scans. I then ran a four-candidate model selection with cross-validation that held out whole days. The decision rule was agreed with the client: a missed faulty run counts 3× a false alarm.
Delivery: Model designs were frozen before building, and predictions were locked before the lab revealed grades. CI and pre-commit guardrails kept model files out of git. A versioned model was released on Hugging Face, with a React prototype and FastAPI model server built by the team. The test set is small, so the final report gives confidence intervals rather than a headline accuracy.
Problem: Deciding whether a release can ship meant reconciling a release brief, GitHub issues, Slack threads and email, about 20 minutes per question. Picking up a stale email gives a customer the wrong date.
What I built: An internal assistant that answers from 6 source types (Slack, email, docs, a live GitHub repo, an exported GitHub snapshot and SQLite) with cited answers. Role-based filtering runs before retrieval, so restricted records never reach the model. A tool-calling agent can only propose actions, and a separate human approval executes them.
Delivery: Thresholds were fixed before testing, and decisions, findings and the threat model were kept in versioned logs. My release call was “demonstrate with explicit limitations”, and the incomplete model comparison is stated rather than hidden.
Problem: Predict departure delay for Tunisair flights, framed for two audiences: the airline operations team and travellers at booking time.
Approach: Benchmarked 5 models on 107,833 flights with time-aware tuning. Error analysis showed the severe-delay tail is driven by knock-on delays from the previous flight, so the traveller-facing product was reframed as a calibrated risk classifier inside a Streamlit “Delay-Alert” app. The 7-day sprint ran on a GitHub project board with 31 issues and 15 PRs, and closed with a retrospective.
Analysed 21,597 home sales pulled from PostgreSQL to advise a luxury buyer persona. Four hypotheses were tested and confirmed. The work ended in a shortlist on interactive maps and a client deck.
21 dashboards across 4 industries — Power BI · Tableau · Looker Studio
Problem solved: Insurance operations teams were manually compiling claims data from multiple sources into static spreadsheets, making live KPI monitoring impossible. This dashboard unified all claims and policy data into a single real-time view, enabling faster decisions on cost reserves and partner commissions.
-- Loss Ratio: claims cost paid as % of earned premium
Loss Ratio (%) =
DIVIDE(
SUM(Claims[ClaimCostPaid]),
SUM(Policies[EarnedPremium]), 0
) * 100
-- Policy MoM Change (Waterfall)
Policy MoM Change =
VAR CurrentMonth = SUM(Policies[ActivePolicies])
VAR PriorMonth =
CALCULATE(
SUM(Policies[ActivePolicies]),
DATEADD(Policies[Date], -1, MONTH)
)
RETURN CurrentMonth - PriorMonth
Problem solved: Renewal teams had no visibility into which policies were at risk of lapsing after repricing. This report surfaced renewal trends by month and policy status, allowing underwriters to prioritise outreach before cancellation windows closed.
Problem solved: Partner data file ingestion was tracked manually via email chains, causing delays in identifying failed or missing submissions. This dashboard automated status monitoring across all partners and source types, significantly reducing data ingestion delays.
Problem solved: When BDX files were flagged for errors, operations teams had no structured way to prioritise remediation. This dashboard ranked flagged files by days overdue and error type, enabling teams to resolve the most critical data issues first.
Problem solved: Aviation finance teams compared airline performance across disconnected spreadsheets with no standard metric definitions. This dashboard consolidated revenue, operating margin, CASK, RASK, and load factor across 5 airlines into one comparable view.
Problem solved: There was no centralised view linking fuel consumption, CO2 intensity, and SAF adoption by region. This dashboard gave sustainability teams a clear picture of where efficiency gains and SAF blending were having the most impact.
Problem solved: Operations managers needed a consistent way to benchmark on-time performance, cancellation rates, and customer satisfaction across multiple airlines. This dashboard standardised KPI definitions and enabled direct peer comparison in one view.
Problem solved: E-commerce managers were spending hours weekly pulling Shopify reports manually and cross-referencing them in Excel. This dashboard connected directly to Shopify data to automate profitability reporting, surfacing order value trends and traffic source ROI at a glance.
Problem solved: The marketing team lacked a unified view of how different user cohorts were converting across devices and channels. By combining Google Analytics session data with revenue metrics, this dashboard helped identify underperforming acquisition channels and informed budget reallocation.
-- New vs Returning user revenue contribution
SELECT
user_type,
COUNT(DISTINCT session_id) AS sessions,
SUM(transaction_revenue) AS total_revenue,
ROUND(
SUM(transactions) * 1.0 /
NULLIF(COUNT(DISTINCT session_id), 0), 4
) AS conversion_rate
FROM analytics_sessions
WHERE date BETWEEN '2023-03-01' AND '2023-04-30'
GROUP BY user_type
ORDER BY total_revenue DESC;
Problem solved: The paid media team managed Google Ads, Meta Ads, and Amazon campaigns in separate platform dashboards, making cross-channel budget comparison nearly impossible. This dashboard consolidated all three into one view, enabling direct ROAS comparison and faster spend optimisation.
Problem solved: Senior marketing leadership had no single report showing cost efficiency (KUR), net revenue margin, and order volume across all channels simultaneously. This replaced a weekly manual slide deck with a live self-service report.
-- KUR: Cost-to-Revenue ratio
KUR (%) =
DIVIDE(SUM(Marketing[Cost]), SUM(Marketing[NettoSales]), 0) * 100
-- CVR: Conversion Rate
CVR (%) =
DIVIDE(SUM(Marketing[Orders]), SUM(Marketing[Sessions]), 0) * 100
-- CPO Netto: Cost per Order
CPO Netto =
DIVIDE(SUM(Marketing[Cost]), SUM(Marketing[Orders]), 0)
Problem solved: Cross-channel performance reviews required pulling from multiple source reports and rebuilding charts each week. This dashboard automated aggregation of sessions, impressions, and KUR KPIs by channel, cutting weekly reporting preparation time significantly.
Problem solved: The SEO and brand team needed a reliable way to track whether brand search volume was growing in response to campaigns. This dashboard tracked Google Search Console impressions and CTR over time, broken down by device, to measure brand visibility gains.
Problem solved: Brand health survey data was delivered as raw CSVs with no ongoing trend view. This dashboard visualised ad awareness and brand awareness across target and non-target groups over time, enabling brand managers to correlate campaign spend with awareness shifts.
Problem solved: The product team lacked a clear view of how revenue from each cohort evolved after the signup date. By combining MoM revenue with cohort analysis, this dashboard helped identify which acquisition periods produced the highest long-term revenue.
-- MoM Revenue Growth
MoM Revenue % =
VAR Curr = SUM(Revenue[Amount])
VAR Prior = CALCULATE(
SUM(Revenue[Amount]),
DATEADD(Revenue[PurchaseDate], -1, MONTH))
RETURN DIVIDE(Curr - Prior, Prior, 0) * 100
-- YoY Revenue Change
YoY Revenue % =
DIVIDE(
SUM(Revenue[Amount]) -
CALCULATE(SUM(Revenue[Amount]),
SAMEPERIODLASTYEAR(Revenue[PurchaseDate])),
CALCULATE(SUM(Revenue[Amount]),
SAMEPERIODLASTYEAR(Revenue[PurchaseDate])), 0
) * 100
Problem solved: The demand generation team had no unified view of how leads progressed from creation through MQL to SQL. This dashboard connected CRM and marketing automation data to give leadership a live funnel view with historical trend analysis.
-- MQL Conversion Rate
MQL Conversion Rate =
DIVIDE(
COUNTROWS(FILTER(Leads, Leads[Stage] = "MQL")),
COUNTROWS(FILTER(Leads, Leads[Stage] = "Marketing Lead")),
0) * 100
-- SQL Conversion Rate (from MQL)
SQL Conversion Rate =
DIVIDE(
COUNTROWS(FILTER(Leads, Leads[Stage] = "SQL")),
COUNTROWS(FILTER(Leads, Leads[Stage] = "MQL")),
0) * 100
Problem solved: Budget vs. performance reporting was done quarterly in static slides with no drill-down capability. This dashboard replaced those slides with a live view comparing planned vs. actual cost and leads by country and channel, enabling in-quarter course corrections.
Problem solved: Field operations teams had no geographic view of where devices were deployed or how they were performing. This dashboard mapped 15,883 devices across Europe with live SQF and EsNo metrics, enabling remote performance monitoring without manual field reports.
Problem solved: Quality engineers needed to quickly identify which devices had the largest deviation between two performance parameters. This dashboard ranked all devices by differential and flagged outliers, replacing a manual spreadsheet review process.
-- SQF Difference between system A and B
SQF Diff = SUM(Devices[SQF_A]) - SUM(Devices[SQF_B])
-- Flag devices outside acceptable range
Out of Range Flag =
IF(
ABS([SQF Diff]) > 10 || ABS([EsNo Diff]) > 5,
"Out of Range", "Within Range"
)
-- Rank by highest SQF Diff
Highest SQF Diff =
RANKX(
ALL(Devices[SerialNumber]),
[SQF Diff], , DESC, DENSE
)
Problem solved: Engineers needed to understand the statistical distribution of device performance to set quality thresholds. The bell-curve visualisation confirmed whether device performance was normally distributed and identified how many devices fell outside the acceptable interval range.
Problem solved: Tracking whether device quality was consistent across a large fleet required comparing two measurement parameters (A and B) simultaneously across thousands of serial numbers. This dashboard enabled engineers to spot systematic deviations at scale.
A scrollable case file for my permission-aware AI agent. It walks through the problem, the threat model, the proof that restricted records never reach the model, the approval gate, and the release decision, using the project's real evaluation charts. The agent itself runs locally in Docker and needs private LLM API keys, so this page presents its evidence rather than a live chat.
Open full screen ↗ Source on GitHub ↗Two maps from my EDA project on 21,597 King County home sales. The first shows the client’s final shortlist: the 4 renovated, Grade 11+ waterfront homes. The second maps 253 luxury homes across the elite zip codes, coloured by size ratio against their neighbours. Hover over a point for price, grade, living area and renovation year. Drag to pan and scroll to zoom.
Map data © OpenStreetMap contributors.
Looking for a BI lead who can also take AI projects from idea to release?