Berke TayfunAkseki.

Problem. Tracking junior roles across Swiss and EU boards by hand meant duplicates, stale posts and no way to rank fit.

Solution. Fetch from RemoteOK and Adzuna, normalise URLs to kill duplicates, score every posting against a candidate profile with TF-IDF + cosine similarity, and serve it all through a JWT-authenticated FastAPI backend in Docker.

[RemoteOK API]   [Adzuna API]
       \             /
   [Ingestion — requests]
              |
   [Dedup — normalised URL hash]
              |
      [TF-IDF matrix] <──> [candidate profile]
              |   cosine similarity → 0–100
        [SQLite] <── [FastAPI + JWT] <── client

Problem. Most portfolio ML is prediction. The experimentation side — deciding whether a launch actually caused a metric to move — is usually skipped.

Solution. Power analysis to size a test, a two-proportion z-test with proper intervals, an SRM check that flags broken randomisation before anyone trusts a p-value, and a DiD estimator for when you can't randomise at all — the estimator family from my thesis.

Problem. Regulatory documents are long and dense; keyword search misses context entirely.

Solution. Ingest PDFs, chunk with overlap, embed into a FAISS index, retrieve the most relevant passages per query, and serve results over REST plus an interactive UI.

Problem. "Compute N-week retention by cohort" and "rank X by Y, handling ties" are easy to get subtly wrong.

Solution. Analytical SQL with no ORM and no pandas standing in for what SQL should do: funnel conversion, weekly cohort retention, rolling averages and country rankings that show the RANK/DENSE_RANK difference on ties.

Problem. The 2025 campaign had mixed results: some stores generated incremental sales, others none — budget was being spent where it didn't work.

Solution. Feature engineering on store data, logistic regression and random forest models to score each store's probability of success, a ranked activation list with an expected-profit "size of prize", and a Power BI dashboard with map-based exploration.

My part. Built the analytics workflow and trained the models in KNIME, supported evaluation (ROC/AUC, confusion matrix) and scoring, contributed to the final delivery.

Problem. A model in a notebook isn't useful to a business stakeholder — it needs to be queryable and visual.

Solution. A preprocessing + training Pipeline evaluated on accuracy, ROC-AUC and a confusion matrix, wrapped behind a prediction API, with a dashboard so a non-technical reviewer can explore predictions directly.

Problem. Real-time multiplayer needs every player to see the same maze, the same enemies and the same damage events.

Solution. The host generates an 18×18 DFS maze at session start and replicates it; NavMesh enemy AI, health and respawn state synchronise through NetworkObject state.

Problem. Prove an understanding of the request/response cycle without the framework's admin doing the work.

Solution. Hand-written models, views and templates; authentication guards on every write route; image upload handling.