Lirose
A job search assistant that reads the offers for you, without making anything up.
- Status
- Live, invite only
- Year
- 2026
- Role
- Product design, architecture and full development
- Stack
- Nuxt
- TypeScript
- Claude
- Zod
- SQLite
- Drizzle
- Telegram
- Typst
Private source code: the architecture and the results are shown here.
The context
Looking for a job means watching a dozen sites, reading duplicate offers and writing the same letters again and again. I wanted to automate this work for several candidates, without ever letting a model invent an experience.
What I built
A deterministic pipeline collects the offers (LinkedIn, Indeed, HelloWork, Welcome to the Jungle, Free-Work, e-mail alerts), removes the duplicates, then has Claude score them with structured outputs validated by Zod. A fast model extracts and scores, a stronger one writes, and six verification layers block any invention. The candidate gets the selection in the app and on Telegram, and applies on their own.
Architecture
- 01CollectJob sites and e-mail alerts, on a regular schedule.
- 02FilterDuplicates removed, hard filters (place, contract, salary).
- 03ScoreExtraction and a 0 to 100 score with structured outputs.
- 04WriteTailored letter and resume, checked against the real profile.
- 05DeliverSelection in the app and on Telegram, the candidate decides.
Results & impact
- About ten sources merged into a single queue, with no duplicates.
- Every offer scored from 0 to 100 against the candidate profile, with the reasons.
- Six anti-invention layers, 54 documented architecture decisions and about 900 tests.
Skills applied
- LLM structured outputs
- Data pipelines
- Full stack product
Got a project in mind?
An application to build, a team to strengthen or an AI product to launch: let's talk.
