Job matching that readsmeaning, not keywords.
Job seekers lose 10+ hours a week to manual searches across LinkedIn, Indeed, and Glassdoor, and keyword matching still misses the roles they are actually qualified for. We built Letti AI: a career platform that scrapes all three, has AI extract 50+ data points from every job, and matches people on real skill overlap with 3072-dimensional embeddings.
Search is manual.Matching is blind.
Job seekers spend countless hours searching across LinkedIn, Indeed, and Glassdoor, applying to hundreds of positions with little sense of whether they are truly qualified or whether the role fits their career trajectory.
Traditional job boards rely on keyword matching, which misses relevant openings whenever a description uses different terminology. Career changers face an even bigger problem: conventional matching algorithms do not recognise transferable skills at all.
Job seekers spend 10+ hours per week searching across multiple platforms, often seeing the same listings repeated.
Traditional matching misses relevant jobs when descriptions use different terminology. 'Full Stack' vs 'Web Developer' vs 'Software Engineer'.
No visibility into salary benchmarks, H1B sponsorship likelihood, or realistic career pivot paths.
Without deep job analysis, candidates cannot tailor their applications or prepare for specific interview questions.
A pipeline that turnslistings into understanding.
Letti AI is a six-stage data pipeline that scrapes, analyses, and matches jobs on semantic understanding rather than keywords. AI models extract 50+ data points per listing, and pgvector with 3072-dimensional embeddings matches people to roles on true skill overlap, so 'React' sits near 'Vue' and 'Data Science' relates to 'Machine Learning'.
Puppeteer with stealth plugins bypasses anti-bot measures to scrape job listings from LinkedIn, Indeed, and Glassdoor. Proxy rotation and request throttling keep collection reliable.
Raw HTML is parsed and sanitised, duplicate jobs are detected and merged, and missing fields are flagged for enrichment in the next stage.
Company data gains additional context: funding stage, company size, industry classification, and Glassdoor ratings when available.
Job descriptions are normalised into a consistent schema, and skills are extracted and mapped to a standardised taxonomy.
AI models extract 50+ data points per job: required and nice-to-have skills, salary hints, remote policy, H1B likelihood, and company culture signals.
Processed jobs are stored in PostgreSQL with pgvector. 3072-dimensional embeddings power semantic search and similarity matching.
What the platformdoes differently.
Every job lands in one of four match bands: Best Match at 80 to 100 percent, Pivot Ready at 60 to 79, Stretch Role at 40 to 59, and Not Ready below 40. Career changers finally see which moves are realistic and which gaps to close first.
Six phases,six months.
From discovery to production deployment in six months.
Discovery and architecture
Analysed job market APIs, designed the scraping architecture, and mapped the skill taxonomy for career matching.
Scraping infrastructure
Built Puppeteer-based scrapers with stealth plugins, proxy rotation, and anti-detection measures for reliable data collection.
AI pipeline development
Developed the AI analysis prompts, embedding generation, pgvector integration, and semantic matching algorithms.
Frontend and UI
Built the React 18 frontend with Radix UI components, career profiling flows, and the interactive job matching interface.
Interview system
Implemented AI interview practice with CV-based question generation, response scoring, and the feedback engine.
Testing and launch
End-to-end testing, performance optimisation, embedding quality validation, and production deployment.