Skip to content

memory.md

Latest commit

 

History

History
203 lines (157 loc) · 10.3 KB

memory.md

File metadata and controls

203 lines (157 loc) · 10.3 KB
Raw

GreenApply — Implementation Progress

Build Status

✅ Builds cleanly — npm run build succeeds with zero errors. Load dist/ as an unpacked extension in Chrome: chrome://extensions → Load unpacked → select dist/.

⚠️ API Key — IMPORTANT

The NVIDIA key shared in chat was exposed. Regenerate it immediately at build.nvidia.com → API Keys. Paste the new key only into the extension: Options → AI Features tab. It stores in chrome.storage.local only — never in any file.

Completed Phases

Phase 1 — Foundation ✅

All project scaffolding in place.

File Purpose
package.json All dependencies declared
vite.config.ts Vite + CRXJS + Tailwind v4
tsconfig.json / tsconfig.node.json TypeScript strict mode
src/manifest.ts MV3 manifest — named platform host permissions, optional *://*/*
.env.example NVIDIA API key placeholder
public/icons/ Green circle placeholder icons (16/48/128px)

Phase 2 — Extraction Pipeline ✅

Heuristic-first extraction — zero API calls in MVP.

File Purpose
src/constants/patterns.ts All regex: languages, visa, work auth, employment type, experience, salary, remote
src/constants/scoring.ts Weights, thresholds, freshness modifiers
src/constants/platforms.ts Per-platform URL patterns + CSS selectors
src/constants/models.ts NVIDIA NIM model names + token budgets
src/background/extraction/jsonld.extractor.ts Parse JobPosting schema
src/background/extraction/regex.extractor.ts Regex extraction for all key fields
src/background/extraction/dict.extractor.ts 80+ skill dictionary matching
src/background/extraction/confidence.ts Per-field confidence scoring
src/background/extraction/pipeline.ts JSON-LD → regex → dict → LLM fallback orchestration

Phase 3 — Resume Parsing ✅

Fully deterministic — no API key required.

File Purpose
src/background/parsers/pdf.parser.ts pdfjs-dist text extraction
src/background/parsers/docx.parser.ts mammoth DOCX text extraction
src/background/parsers/resume.parser.ts Deterministic ResumeProfile parser with date range merging

Key implementation: computeMergedExperienceYears() unions overlapping date intervals before summing — prevents inflated experience from concurrent roles.

Phase 4 — IndexedDB Stores ✅

Store Key Purpose
profile 'main' Single UserProfile record
jobs jobId Raw JobListing scraped data
extractions jobId Cached ExtractionResult — revisit = instant
matches jobId Cached MatchResult — revisit = instant
applications id Full CRUD with status timeline
companies normalizedName CompanyProfile schema (seeded post-MVP)
rules id User automation rules
metrics metric key Incrementing usage counters

Phase 5 — Scoring Engine ✅

File Purpose
src/background/scoring/hard-filters.ts 6 hard filters: language gap, visa, work auth, employment type, location, experience gap, excluded company
src/background/scoring/engine.ts Deterministic 0–100 scorer — 7 weighted dimensions + freshness modifier
src/background/scoring/freshness.ts Age-based score modifier (+5 fresh, -5 stale, -15 very stale)
src/background/rules/engine.ts Evaluate UserRule conditions against match + extraction
src/background/rules/defaults.ts Auto-generate default rules from workAuth status

Score weights: skills(30%), experience(22%), language(20%), location(10%), visa(8%), salary(5%), employment type(5%).

Hard filter types: language_gap, visa_blocked, location_blocked, employment_type_mismatch, experience_gap, excluded_company. Any blocker severity forces score ≤ 34 and recommendation = red.

Phase 6 — Background Handlers ✅

File Purpose
src/background/handlers/job.handler.ts Fingerprint job, deduplicate, run extraction pipeline, cache
src/background/handlers/match.handler.ts Cache-check, compute match, push result to tab, update metrics
src/background/handlers/profile.handler.ts Save profile, upload + parse resume, seed default rules
src/background/handlers/tracker.handler.ts Save application, update status, increment metrics
src/background/index.ts Service worker message router — handles all BackgroundMessage types
src/shared/utils/fingerprint.ts fingerprintJob() — sha256 of normalized company+title+location

Job fingerprinting: normalizeTitle() strips (m/f/d), (w/m/d), senior, junior etc. before hashing — prevents duplicate jobs from different URLs counting separately.

Phase 7 — Content Scripts ✅

File Purpose
src/content/observer.ts History patch (pushState/replaceState) + MutationObserver — detects SPA navigation
src/content/detectors/generic.detector.ts PRIMARY — JSON-LD + URL heuristic + DOM keyword scan
src/content/detectors/linkedin.detector.ts LinkedIn-specific URL pattern
src/content/extractors/generic.extractor.ts JSON-LD → microdata → largest text block
src/content/extractors/linkedin.extractor.ts LinkedIn-specific CSS selectors (aria-stable attrs)
src/content/shadow-host.ts Closed shadow DOM mount — CSP-safe, style isolation
src/content/index.ts Entry — detect → extract → mount overlay → send JOB_DETECTED
src/content/overlay-mount.tsx React root inside shadow DOM

Phase 8 — Overlay UI ✅

Component Purpose
Overlay.tsx Main container — listens for greenapply:message events, state machine
ScoreRing.tsx Animated SVG progress ring (0–100)
RecommendationBadge.tsx 🟢🟡🟠🔴 pill with label
HardFilterAlert.tsx Red/orange alert cards for blockers and warnings
SkillGapList.tsx Matched (green) / bonus (blue) / missing (red) skill chips
JobFreshness.tsx "Posted 87 days ago" with warning for stale listings
ConfidenceCaveat.tsx Low-confidence visa notice
ActionButtons.tsx "Save Job" button → sends SAVE_APPLICATION to background

Core UI moments implemented:

  • 🔴 Skip with "Estimated time saved: ~15 minutes" badge
  • 🟢 Strong Apply with score ring + skill breakdown + save button

Phase 9 — Popup + Options ✅

File Purpose
src/popup/Popup.tsx Dashboard (time saved, stats) + Jobs list with status badges
src/options/Options.tsx 4 sections: Resume upload, Profile, Preferences, API key (optional)

Phase 10 — AI Layer + Remaining Features ✅ (added 2026-05-30)

File Purpose
src/background/nim/client.ts nimComplete() (batch) + nimStream() (SSE) with 3-attempt exponential backoff
src/background/nim/explainer.ts 2–3 sentence score explanation — called async after match computed, result pushed to tab
src/background/nim/generator.ts Cover letter streaming prompt — candidate profile + job desc
src/background/handlers/generate.handler.ts Streaming via chrome.runtime.connect() port — keeps SW alive
src/overlay/components/GeneratePanel.tsx Streaming cover letter panel with copy button
src/overlay/components/TrackingDropdown.tsx Status picker (saved → applied → interview → offer → rejected)
src/content/feed.ts Injects 🟢🔴 score badges into job list cards (LinkedIn, Indeed, StepStone)

New platform extractors + detectors: Indeed, Greenhouse, Lever, Workday, Personio, StepStone — all wired into content/index.ts.

How AI works:

  • Match score is computed deterministically first and returned immediately
  • LLM explanation fires async in background; when done, pushes updated MatchResult to the tab (overlay re-renders with summary text)
  • Cover letter streams token-by-token through a long-lived chrome.runtime.connect() port
  • If no API key: score + hard filters still work; explanation and cover letter are hidden
  • Extraction is now AI-first for testing: the full job text goes to NIM first, then JSON-LD / regex / dictionary fall back for missing fields.

Next Steps (not yet built)

Extraction quality gate (before shipping)

Test on 100 real job pages. Targets:

  • Language detection >90%
  • Visa/sponsorship >85% accuracy, <5% false positives
  • Employment type >90%
  • Experience years >80%

Phase 10 — Feed Filtering

src/content/feed.ts — inject 🟢🔴 score badges directly into job list cards on LinkedIn / Indeed before user clicks through.

Phase 11 — NIM AI features (optional, API key required)

  • src/background/nim/client.ts — fetch wrapper
  • src/background/nim/explainer.ts — 2–3 sentence score summary
  • src/background/nim/generator.ts — streaming cover letter via chrome.runtime.connect() port
  • src/overlay/components/GeneratePanel.tsx — streaming text output

Phase 12 — Additional platform extractors

Indeed, Glassdoor, StepStone, Workday, Personio, Greenhouse, Lever, Ashby.

Architecture Notes

  • No API key required for MVP. All scoring and extraction is deterministic.
  • Extraction cache: extractions store keyed by jobId — revisiting same job skips all processing.
  • Match cache: matches store keyed by jobId — re-opening same job tab = instant result.
  • Shadow DOM overlay: Closed shadow root prevents page CSS from leaking in. Tailwind styles injected via adoptedStyleSheets — CSP-safe.
  • Chrome Store strategy: Named platform domains in content_scripts.matches, *://*/* only in optional_host_permissions. Generic career sites request origin permission at runtime.
  • Date range merging: computeMergedExperienceYears() unions overlapping intervals — critical for accurate seniority inference on student profiles.
  • Feature updates: When adding or changing a feature, update memory.md with the new behavior or convention.

How to Load in Chrome

npm install
npm run build
# Chrome → chrome://extensions → Enable Developer mode
# Load unpacked → select /Users/khaleel/Documents/greenapply/dist/

Visit any LinkedIn job page — the overlay should appear bottom-right. Upload resume in the extension options page first for match scoring to work.