Staff Frontend Engineer

& Technical Leader

CV tools are either pretty and unparseable or parseable and ugly. Both sell the same promise — "ATS-friendly" — and neither can show you what an automated screen actually read back.

A PDF that looks correct proves nothing: the text layer can arrive out of reading order, or be an image with no text at all. And the render path is WASM, which a green build silently drops — the bundler cannot see a file read at runtime, so dev and CI both pass while production has no renderer.

A training log is health data, and the box where you describe your situation is where somebody writes about an injury. Every tracker asks for an account first and holds that in plain text on its own server.

Encrypting on the device is the easy half. One encrypted blob per account is fine on one device and destroys data on two: train on the phone, open the laptop holding a stale copy, and the session is gone. Syncing without ever letting the server read a workout means the merge has to work on data it cannot see.

A corpus of 927 product plans, five documents each, sitting in folders. Deciding what to build next meant opening directories one at a time, and a corpus you cannot search is a corpus you rewrite by accident.

The facets a reader wants are buried in prose written by successive generations of tooling: the country is a loose line, what someone will pay is free text, the stack is a bold label halfway down a plan. Faceted search over that means parsing prose reliably — and the whole thing had to keep running on a static file server.

Hiring and open-source sourcing both run on GitHub stars, which reward what was popular years ago rather than who is shipping this month.

Thirteen platforms expose activity in thirteen shapes, the same person appears under different handles on each, and every signal ages at a different rate. Ranking them together means reconciling identity and freshness before a single result can be trusted.

Finding anything in Copenhagen means four different sites: Boliga for flats, DBA for cars, Facebook groups for services, Airbnb for the rest. Each one is a feed, and a feed cannot tell you what is ten minutes from your door.

Properties, vehicles, services and experiences have almost nothing in common as data: different prices, different units, different lifespans. Putting all four on a single map means one schema flexible enough to hold them and one interface that does not turn into a filter panel.

Every AI CLI keeps its own config, its own skills folder and its own MCP list. Six of them means six copies of the same setup, drifting apart from the day you install the second one.

A dev environment is not just dotfiles: it is CLIs, runtimes, language servers, model endpoints and the prompts themselves, spread across macOS, Linux and Windows. Making it reproducible means it has to be safe to run twice on a machine that is already half configured.

shadcn/ui gives you every primitive a chat interface needs and no chat interface. Everyone rebuilds the same streaming panel, and everyone rebuilds it welded to one provider.

Streaming, markdown, code blocks, file upload and virtualised history are each easy alone and awkward together, especially while tokens are still arriving. Doing it once for Anthropic and again for every OpenAI-compatible gateway would have meant maintaining the same panel eight times.