Skip to content

Latest commit

 

History

History
165 lines (129 loc) · 6.5 KB

File metadata and controls

165 lines (129 loc) · 6.5 KB

César Roadmap

This roadmap collects all ideas discussed during the initial build sessions and community feedback. Items are organized by theme and priority. Contributions welcome on any item.

Status legend

  • ✅ Done
  • 🔨 In progress
  • 📋 Planned
  • 💡 Idea (needs discussion)

Detection engine

✅ Shipped

  • Two-axis detection model (third-party + gating CTA)
  • 3-layer confidence scoring (regex 50% → behavioral 75% → LLM 99%)
  • Comment bait analysis (emoji/keyword ratio in comments section)
  • Public URL detection (GitHub repos, npm packages, official docs)
  • Author-product mismatch (author doesn't work at mentioned company)
  • Post structure fingerprint (hype intro → feature list → CTA at end)
  • Auto-expand "voir plus" / "see more" for full text scanning
  • MutationObserver re-scan when user manually expands a post

📋 Planned

  • Cross-reference author LinkedIn profile with mentioned products (deeper than subtitle matching)
  • Detect carousel/PDF posts where the gating is in the image, not the text
  • Scan shared/reposted content (currently only scans original posts)
  • Detect "link in first comment" patterns (scan comments for author's own link)
  • Language expansion: DE, ES, PT, NL, IT regex patterns
  • Detect gating via LinkedIn newsletters ("subscribe to get...")

💡 Ideas

  • Image OCR on carousels to detect CTA text in images
  • Detect "fake scarcity" patterns ("only 50 spots left", "I'm removing this tomorrow")
  • Track post edit history — some authors add CTA after the post gains traction
  • Sentiment analysis on comments to distinguish genuine engagement from bait responses
  • Detect engagement pods (same accounts always commenting first)

LLM integration

✅ Shipped

  • Multi-LLM support: Anthropic (Haiku 4.5), OpenAI (GPT-5 mini), Google (Gemini 3 Flash), xAI (Grok 4.1 Fast)
  • Auto-detect provider from API key prefix
  • Prompt caching for Anthropic (cache_control: ephemeral)
  • 4-strategy JSON parser with truncation repair
  • Language-adaptive output (LLM responds in post language)
  • Cleared badge when LLM determines post is legitimate
  • Strengthened prompt to avoid false positives on original content

📋 Planned

  • Web search integration (LLM uses web search to verify source URLs)
  • Confidence calibration — compare LLM confidence with actual user feedback
  • Prompt A/B testing framework to optimize false positive rate
  • Support for local/self-hosted models (Ollama, LM Studio)

💡 Ideas

  • Multi-LLM consensus — query 2 providers, flag only if both agree
  • Fine-tuned micro-model specifically for parasitic detection (distilled from labeled data)
  • LLM-powered "source finder" that returns the exact original URL

User interface

✅ Shipped

  • Dark blue badge with 3 states: pending (orange), flagged (red), cleared (green border)
  • Reason text on its own line below header (no truncation)
  • Collapsible comment section (purple toggle)
  • Comment pre-fill in LinkedIn's comment box
  • "Detected by César 🏛️" signature in generated comments
  • Feedback buttons: Confirm / Partial / False positive
  • Auto-fade on cleared badges (8s)
  • Minimal popup with reveal settings panel
  • API key guard on web search toggle

📋 Planned

  • Badge position preference (top vs bottom of post)
  • "Why was this flagged?" expandable explanation with layer breakdown
  • Dark/light theme toggle for the badge
  • Keyboard shortcuts (dismiss badge, copy comment, etc.)
  • Badge counter in extension icon (like unread messages)

💡 Ideas

  • Floating dashboard showing detection stats across sessions
  • "César score" displayed next to author name on all their posts
  • Mini-tutorial overlay on first install
  • Customizable comment templates (user writes their own tone)
  • Right-click context menu: "Analyze this post with César"

Community & backend

📋 Planned (high priority)

  • Backend API + database (Supabase) for storing detections
  • Shared flagging: if 3+ users flag the same post, instant detection for all
  • Recidivist leaderboard: most flagged accounts across all users
  • Community-submitted source directory: "this post gates → this URL"
  • Vote system: confirmed / false positive (improves scoring over time)
  • Anonymous usage analytics (opt-in) to improve detection patterns

💡 Ideas

  • Public API for researchers studying engagement bait on LinkedIn
  • Browser-based community dashboard (like a subreddit for flagged posts)
  • "César verified" badge for accounts that never gate others' content
  • Integration with LinkedIn's own reporting system
  • Webhook notifications when a recidivist posts new content
  • Gamification: "sourceur" leaderboard for users who submit the most sources

Distribution & growth

✅ Shipped

  • GitHub repository (public, open source)
  • Chrome Web Store publication (privacy policy, store listing, packaged build)

📋 Planned

  • LinkedIn announcement post (the growth loop)
  • Landing page explaining the project

💡 Ideas

  • Firefox support (WebExtension API — minimal changes needed)
  • Edge Add-ons store publication
  • Safari support (longer term, requires Xcode wrapper)
  • "Powered by César" watermark on shared screenshots
  • Referral system: "installed by X users from your network"
  • Partnership with LinkedIn anti-spam/creator integrity initiatives
  • Conference talks / meetup presentations on engagement bait detection
  • Open dataset of labeled parasitic vs legitimate lead magnets

Technical debt & quality

📋 Planned

  • Comprehensive test suite (synthetic posts for each detection pattern)
  • LinkedIn DOM selector monitoring (they change frequently)
  • Error reporting system (opt-in, captures PARSE_ERROR and selector failures)
  • Performance profiling (scan time per post, LLM latency)
  • TypeScript migration (type safety on detection engine)

💡 Ideas

  • Automated visual regression testing on badge UI
  • CI/CD pipeline with Chrome extension packaging
  • Canary deployment (test new patterns on 10% of users before full rollout)
  • Extension size budget monitoring
  • Memory leak detection (long-running MutationObserver)

Contributing

Pick any item and open an issue or PR. The most impactful areas for new contributors are:

  1. Regex patterns — add patterns for new languages or edge cases
  2. LinkedIn DOM selectors — these break frequently and need community monitoring
  3. LLM prompt tuning — help reduce false positives with real-world examples
  4. Testing — scroll your feed with debug mode on and report what César gets right and wrong

See CONTRIBUTING.md for how to get started.