English
Video coming soon — drop
assets/intro-en.mp4 into this project.
Quick introductions
About a minute each. Same person, two languages — useful if you are hiring in Switzerland or an English-speaking team.
Deutsch
Video folgt — Datei
assets/intro-de.mp4 hier ablegen.
Projects & technical portfolio
A few things I’ve built — systems work, cloud, security, and places where I used AI because it genuinely helped.
Professional case study
SentinelHub
I designed and built this from scratch for my former employer: a platform that watches internet-facing systems and shows brokers clear cyber evidence — without pretending to make underwriting decisions for them.
- Why I built it
- Brokers needed ongoing proof of a company’s security posture, not a one-off spreadsheet. I wanted checks that run the same way every time, plus a place to see findings and history.
- What I made
- The full product: Angular UI, Go API, MongoDB, Auth0 login, a scheduler for passive checks (TLS, DNS, headers, health, and similar), plus controls, benchmarks, findings and reports. They kept it for further development after I left.
- Where I used AI — and where I didn’t
- I used AI tools heavily while coding and designing, because that let me move faster on a large build. The monitoring itself is normal code: predictable checks, not a model guessing. For broker assessment write-ups I did add a constrained LLM step — the facts stay structured and verified in code; the model only helps turn them into a readable report, which is something prompts are better at than hand-writing templates for every case.
A bit more detail
- Evidence first, then humans decide risk appetite — SentinelHub doesn’t try to be the underwriter.
- Core pieces: assets, targets, checks, results, companies, controls, benchmarks, findings and reports.
- Runs as Docker Compose; the scheduler lives in the API process (no Kubernetes).
Personal · privacy-focused
Home ETL — personal finance assistant
I built this for our household so bank PDFs stay on our own server. It reads statements, keeps letters separate, and shows me honestly where months of history are missing.
- Why I built it
- I didn’t want our finances living only in someone else’s cloud. I also got tired of guessing whether I’d actually uploaded every statement.
- What I made
- A small self-hosted stack: upload PDFs, extract what I can, let me approve anything before it counts as truth, then analyse spending. It currently understands Yuh, UBS (including scanned pages via OCR), and Santander UK.
- Where I used AI — and where I didn’t
- Bank statement layouts are predictable, so I wrote normal parsers in code for those. Scanned UBS PDFs need OCR (Tesseract) — that’s a tool, not a chat prompt. I don’t use an LLM to invent balances: numbers only become “real” after I review them. AI helped me build and iterate the app itself; it doesn’t get to rewrite our ledger.
Spending analytics
I wanted charts I could trust. If a month has no statement, I leave it blank instead of drawing a fake zero — otherwise averages lie. The table says “no statement covers this month” when that’s the case.
Letters vs statements
Banks also send rate-change letters and product notices. Those aren’t transactions, so I route them to Mail instead of the ledger. Code classifies the PDF type; I keep the summary and the original file, but nothing here changes balances.
Missing statement coverage
This grid answers a simple question: which months am I missing? Green means I have a statement; red means there’s a gap between uploads I already made. If OCR couldn’t read a file cleanly, I flag it for review — having the PDF is not the same as trusting the numbers.
A bit more detail
- Postgres stores what I’ve approved; original PDFs stay in MinIO as evidence.
- Flow: upload → extract → classify → parse → I review → promote.
- Partner vs accountant roles so access stays limited.
Planned
Next: a smaller AI ops project
I still want a compact public project that shows AI helping with day-to-day IT work — not another big private app.
- What I’m thinking about
- Something practical: maybe log triage, drafting runbooks, or automating a boring admin task. I’ll pick one, build it, and replace this card when it’s real.
A bit more detail
TODO for me: choose the project, decide if the repo will be public, then rewrite this card.