Matchra

I was spending 20–30 minutes per job application just deciding if a role was worth pursuing.

Reading JDs carefully, cross-referencing my resume, second-guessing myself. Half the time I'd write a full cover letter for something I was underqualified for, or skip a role I was actually a strong fit for because the title threw me off. I wanted a second opinion I could get in under a minute, from my phone, without uploading my resume to yet another platform.

Instead of building a web app, I built a Telegram bot.

Because that's where I'm already spending time during a job search. The whole system runs locally: n8n orchestrates the workflow, Ollama runs the LLM on-device, and nothing leaves your machine. The bot accepts your resume once, then you send job descriptions or links whenever you find a role. It returns a score, a breakdown, and a plain-language explanation in seconds.

The scoring is deterministic.

A separate engine computes the match score from structured data the LLM extracts, rather than asking the LLM to score directly. This was a deliberate tradeoff: it means the same resume against the same JD always produces the same number, but it also means the score is only as good as the extraction quality. The LLM's role is narrowed to normalization and explanation — two things it does reliably — while the math stays in code where it can be audited and adjusted.