Job Focus Assist

Local Chrome tracer for job hunting on a MacBook Air M4 (24 GB). It records pages you already opened, keeps a should-apply / continue / applied / pending pipeline on this Mac, and only then (if you ask) scores a posting against your resume with a local model.

It does not crawl career sites, click Apply, or message anyone. Opening a job is not Analyze.

Pipeline

seen → should apply → continue → applied → pending → completed
                              ↘ skipped

History is an event log (visited, imported, scored, submitted, status changed). The current card is the job row.

What runs the local model

Suggestions have to name a resume line. Gaps get a substitute (Kubernetes → Docker at Qure.ai; Go → Java/Spring + C++ fibers) or a hard-blocker label. Empty shrugs are dropped.

ATS boolean filters are a separate check (no model): if the JD says React.js and the CV says React, that is a spelling trap, not a missing skill. The tracker lists the exact string to put on the CV.

Why these models

On this fanless 24 GB Air, Chrome + LinkedIn already use a lot of RAM. Marketing posts that say “24 GB runs Qwen 3.6 27B” assume Chrome is closed.

Thinking mode is off. Context is capped at 8k. Ollama unloads after 5 minutes idle.

Storage

• Settings and profile: chrome.storage.local
• Job history: IndexedDB jfa in the extension origin (not cookies, not sync, not a server)
• Existing inbox rows migrate once (new → seen, later → should apply)

Install

1. Local model

chmod +x ~/job-focus-assist/scripts/setup-ollama.sh
~/job-focus-assist/scripts/setup-ollama.sh

Or by hand:

brew install ollama
launchctl setenv OLLAMA_ORIGINS "chrome-extension://*"
launchctl setenv OLLAMA_KEEP_ALIVE "5m"
launchctl setenv OLLAMA_NUM_CTX "8192"
open -a Ollama
ollama pull qwen3.5:9b
ollama pull qwen3.5:4b

Restart Chrome after setting OLLAMA_ORIGINS so the extension can reach 127.0.0.1:11434.

2. Load the extension

1. Chrome → chrome://extensions → enable Developer mode
2. Load unpacked → select ~/job-focus-assist (the folder that now contains manifest.json)
3. Pin the toolbar icon
4. Turn it On in the popup

3. Try the demos (no login needed)

Open extension/demo/sample-job.html and click Analyze this tab.
Open extension/demo/sample-microsoft.html — the visit is saved as Seen with Eightfold ids 1970393556957799 / 200047330. The #submitted hash is a confirmation page and logs Applied.
Open extension/demo/sample-meesho.html for the Lever flow (Meesho SDE II Data): job → #apply (Continue) → #submitted (Applied). The apply and thanks URLs collapse to the same posting id.
Open extension/demo/sample-keka.html for Keka Hire (Scry AI Java Developer): visit → Seen, #submitted → Applied.

If you want the on-page chip on file:// pages, enable Allow access to file URLs on the extension’s details page.

How to use it

1. Turn the extension On. Off means it does nothing.
2. Open a job detail page. The board is detected from the ATS (Lever, Greenhouse, Keka, Eightfold, Ashby, Workday, LinkedIn, Amazon) — Meesho is Lever, Microsoft is Eightfold, Scry AI is Keka. Visit is stored. No model. /apply is Continue. “Application submitted!” is Applied.
3. On the chip: Should apply, Continue later, Mark applied, or Analyze.
4. Open the tracker side panel. Filter To do / In progress / Waiting / Done.
5. On a search list, Score visible — only cards on screen, no pagination.
6. Microsoft saved-jobs JSON that the page already loaded is imported (we do not call Eightfold ourselves).
7. On any other careers page, popup → Analyze this tab.

A daily reminder (default 20:00) fires for Should apply and Continue with score ≥ 65. Glanced-at Seen rows do not nag.

What it looks for (your 2026 profile)

Seeded from your resume: IIT (ISM) Dhanbad CSE 8.80, Amazon SDE intern (Hyderabad), Qure.ai AI intern (Bengaluru), InferMesh, fiber scheduler, io_uring server.

• New-grad / SDE / backend / systems / ML infra titles
• Stack you can prove: Java/Spring, Kafka, SQS, Python/Django, C++, Docker, AWS, Prometheus, inference
• Red flags: citizenship, clearance, 4+ years, staff/principal, people-management
• Edit all of this in Options

Privacy and limits

• Job text and your profile never leave this Mac (Ollama on localhost).
• No <all_urls> crawl permission. Content scripts are host-allowlisted.
• We never request Microsoft/Eightfold APIs. If you are logged in and the page already fetched saved jobs, we may read that response.
• LinkedIn DOM changes. If an extractor misses, use Analyze this tab.
• Do not leave 15 other heavy extensions + 9B + 40 LinkedIn tabs running.

Project layout

extension/     Chrome MV3
  background/  Tracer, Ollama (suggestions only), reminders
  content/     Chip, extractors, Microsoft page hook
  lib/         IndexedDB, pipeline, resume alternatives
  popup/       On/Off + health
  sidepanel/   Tracker + history
  options/     Profile + model
  demo/        Offline sample postings
scripts/       Icon generator + Ollama setup

Checks

node scripts/test-match.mjs