Statistics student in Istanbul who builds AI-assisted systems that collect data, check it for errors, and turn it into reports.
Looking for: operations or data/analytics roles at early-stage startups. 3.5 days a week during term (Istanbul or remote), full time after June 2027.
| Project | Business problem | Result you can check |
|---|---|---|
| MizanKöprü | Month-end consolidation and internal control for a group of companies reporting in several currencies | Flags all 14 errors deliberately planted in its synthetic demo data; 24 internal control tests |
| RafRadar | How visible is a brand on the online shelf, meaning Trendyol and N11 search results? | Bundled demo: 24 searches, 1,222 listings. The "multivitamin" search returns 36 results, none from the brand |
| LeadRadar | Find local businesses with missing or broken websites and qualify them as sales leads | One PDF page of evidence per prospect; a HIGH score needs verified findings, impressions alone never qualify |
| gozetimli-ogrenme, gozetimsiz-ogrenme | Is a model's reported accuracy real? Reference code written from scratch and checked against scikit-learn | 562 and 275 tests. Rebuilding every worked example exposed 5 and 10 errors in the course notes |
| Angie | A personal assistant in Turkish over Telegram: calendar, mail, tasks, contacts | Runs on n8n 2.x, where its template broke; documents 10 n8n pitfalls and fixes 4 template bugs |
| Bölüm Takip | Tracks master's applications around the deadline that actually binds a non-EU applicant | Deployed on Vercel as a login-protected personal tool; 240 automated checks |
- The engine produces the numbers; AI only interprets them. In MizanKöprü, switching the AI layer off leaves every figure unchanged.
- Missing data is never skipped silently. A step that cannot run is reported with its reason, and the closing verdict becomes conditional.
- Every system has an independent check. MizanKöprü re-reads the raw files without the pipeline's own code; its demo data carries an answer key the pipeline never reads.
Stack: Python, pandas, NumPy, scikit-learn, TypeScript, Next.js, Postgres, n8n, Claude API.
Most repositories are source-available: free for noncommercial use (PolyForm Noncommercial 1.0.0); commercial use requires a licence.