Cross-user search across every resume / cover-letter PDF the platform has ever generated — hot, trashed, archived, and policy-purged. Downloads + restores write to the impersonation audit log.

UserAny Company Title State Anyhottrashedarchivedpurged KindEitherResumeCover letter Fit min Fit max Search Reset

This UI is grounded in the visual language of Luc Besson's The Fifth Element (1997) — specifically Jean-Paul Gaultier's costume design and the film's retro-futurist world-building. The goal is a lived-in, high-density future: not sterile sci-fi white, but layered, tactile, and slightly worn.

Core Palette

Components

Typography

Space Grotesk — angular, geometric letterforms for all UI chrome, headings, labels, and badge text. Draws from the film's utilitarian industrial stencil language.
IBM Plex Mono — all data output, terminal logs, scores, paths, and thinking traces. The machine's voice.

Structural Language

Panels use clip-path chamfered corners — the same angular, armor-panel geometry seen in Gaultier's bandage-wrap costumes and the film's industrial hardware. Every surface suggests function over decoration.

Scanline overlay references the film's omnipresent screen culture — the city is surveilled, everything is displayed, nothing is private.

Reference Material

Design Rules

• Orange (#FF6B1A) is active / interactive. Use sparingly — primary actions only. When it appears, it commands.
• Green (#39FF14) is success / confirmed / safe. Terminal log text, pipeline completion, submitted state.
• Yellow (#FFD700) is pending / requires review. Awaiting human action.
• Red (#C0392B) is error / blocked / rejected. Quieter than magenta — deliberate, not panicked.
• No rounded corners on structural elements — chamfered only. Softness contradicts the aesthetic.
• All UI chrome uppercase or small-caps. Mixed case only in user-generated content (job titles, descriptions).
• Glow effects are sparse — used to signal state (online, active, hover), never decoration.
• Dense information layout — the city never has empty space. Panels should feel like instrument clusters.
• Borders are structural, not decorative. They define the shape, like Gaultier's bandages define the body.

Job Gobbler is a self-hosted job application automation service running on port 7432. It is orchestrated by n8n (4 workflows) and surfaces results via Telegram. The pipeline discovers LinkedIn job postings, scores them with a local LLM, and prepares tailored resumes — all without sharing data with third parties (except OpenRouter for resume writing).

Starting the service:

Listens on http://localhost:7432. n8n reaches it via host.docker.internal:7432.

1. Search

   n8n workflow 01 fires every 15 minutes and POSTs each search URL to /jobs/search sequentially. Playwright opens LinkedIn with your saved profile, scrolls the job list to hydrate all cards, and stores up to 50 new job IDs in SQLite.

2. Describe

   /jobs/describe fetches each job's detail page via plain HTTPS (no browser, no auth needed). Parses title, company, salary, location, description text, applicant count, Easy Apply flag. Stored in description_text.

3. Grade

   /jobs/grade-all scores all described jobs via a single-shot call to Qwen3-Coder-30B (local, ~800ms/job). Produces fitScore, scoreCause, riskFlags, and per-dimension scores.

4. Telegram

   Graded results are formatted and sent to your Telegram chat. Approval buttons let you reject, prep resume, or proceed to apply.

All inference is local except resume/cover letter writing.

Task	Model	Location
job_ranking · fit_scoring · risk_flags	Qwen3-Coder-30B	LOCAL
form_extraction · structured_output	Qwen3-Coder-30B	LOCAL
visual_qa · screenshot_verification	Gemma 4 27B	LOCAL
resume_generation · cover_letter	Gemini 2.5 Flash → DeepSeek fallback	REMOTE

LM Studio endpoint: http://192.168.0.14:1234/v1

ID	Name	Trigger
01	Discovery Scheduler	Every 15 min → search → describe → grade
02	Telegram Approval Router	Telegram callback_data → status/apply/build
03	Direct Job Intake	Telegram URL drop → scrape → grade → notify
04	Health Monitor	Periodic → GET /health → alert if down

Push changes: cd n8n-workflows && ./sync.sh push

Method	Path	Description
GET	/health	Service status, policy list, routing stats, audit summary
GET	/pool	Browser profile pool — free/busy/queued
GET	/audit	Recent LLM call log (filter by model, task, jobId)
GET	/current-tab?url=	Firefox bridge — scrape + grade the active tab URL
POST	/jobs/search	LinkedIn discovery — returns new job IDs stored in DB
POST	/jobs/describe	Headless fetch of job description pages (no auth)
POST	/jobs/grade-all	Single-shot 6-dim scoring (~800ms/job) via Qwen3-Coder-30B
POST	/jobs/scrape	Scrape any job URL via browser
POST	/jobs/notify	Format graded jobs into Telegram message payloads
POST	/jobs/build	Generate tailored resume + cover letter (PDF)
POST	/jobs/apply	Initiate apply pipeline — routes by detected platform
POST	/jobs/fill	LinkedIn Easy Apply form-filler (returns plan for approval)
POST	/jobs/fill/approve	Execute approved fill plan (submit button)
POST	/jobs/status	Update job status (rejected / approved / applied)

SQLite at data/jobs.db. Key columns:

Status flow: new → described → graded → approved / rejected / applied