JustHireMe

JustHireMe is a desktop workbench for running a job search on your own machine. It collects postings from many sources, filters and ranks them against your profile, and generates a tailored application package for the roles worth pursuing. Profile data, lead history, generated documents, graph data, vectors, and settings stay local by default.

The pipeline runs in stages. Source adapters pull leads from applicant-tracking and company boards, RSS and Atom feeds, Hacker News, GitHub, Reddit and community sources, custom JSON endpoints, and a browser-backed web fallback; a scout module fans out across them and deduplicates. A deterministic quality gate then rejects or penalizes postings with no apply URL, thin scraped text, stale dates, senior-only seniority in beginner-focused feeds, unpaid or commission-only red flags, or missing company context, recording the reason for each rejection.

Ranking combines source signal, lead quality score, a deterministic fit rubric, seniority caps, project and stack evidence, semantic similarity from local vectors, feedback learning, and optional LLM evaluation, and each fit score is backed by evidence from your own skill and project graph. For a strong lead the customizer produces a tailored resume PDF, a cover letter PDF, a founder message, a LinkedIn note, a cold email, a keyword-coverage summary, and the rationale for which projects were selected. Scoring is relative to the candidate's own domain rather than a fixed technology vocabulary, so discovery and tailoring work across fields, and the target location is detected from the resume or set explicitly.

The application is a Tauri 2 desktop shell around a React 19 and TypeScript front end and a Python 3.13 FastAPI sidecar with WebSockets. Local storage uses SQLite for the lead CRM, Kuzu for the profile graph, and LanceDB for vectors; embeddings come from a bundled ONNX all-MiniLM-L6-v2 model with a deterministic hashing fallback, so semantic matching runs offline with no API key. Discovery, ranking, and generation can run keyless through Ollama or an existing Claude Code or Codex CLI subscription, with keyed providers such as OpenAI, Anthropic, Gemini, Groq, and DeepSeek as options. Browser automation and auto-apply exist as an experimental lab area that is disabled by default and outside the supported core.

Two agent surfaces ship with the project: an agent-neutral skill file that tells a coding assistant how to work inside the repository, and a stdio MCP server exposing score_job_fit, evaluate_lead_quality, and extract_lead_intel.