FixMyFeed

FixMyFeed is a behavior-aware social media filter built around one simple idea:

Jayden had much lower social-media screen time than the rest of the team, and he credited one habit above all else: keeping his phone in grayscale.
FixMyFeed takes that insight and turns it into a full attention-shaping system:

• AI triages reels in real time (SKIP, WAIT, LIKE_AND_STAY)
• the feed can be auto-scrolled/locked based on that decision
• color is treated as a limited credit budget (grayscale as friction)
• behavior data is logged and analyzed into dashboards and coaching insights

• Chrome extension (extension/)

  • Popup settings UI (interests, blocked keywords, autolike, user id)
  • Content script that evaluates reels, blocks/skips content, and manages overlay UX
  • Background service worker that handles API calls, user identity, credit state, and parental sync

• Backend API (main.py)

  • FastAPI orchestration service
  • Multi-agent decision + analysis pipeline through Lava-hosted model endpoints
  • Supabase persistence and analytics endpoints

• Dashboard web app (dashboard/)

  • Neural map / analytics / AI insights / parental controls
  • Visual feedback for rewiring progress and content consumption patterns

1. Extension content script captures visible reel text.
2. Background worker sends /evaluate request with user interests + avoid keywords.
3. Backend gatekeeper returns one of:
   • SKIP
   • WAIT
   • LIKE_AND_STAY
4. Content script applies behavior:
   • SKIP -> keep blocker, auto-scroll
   • WAIT -> currently behaves like stay (no auto-like)
   • LIKE_AND_STAY -> stay and optionally like
5. Watch events are logged via /log_watch.
6. Dashboard pulls /stats and /insights for user-level reporting.

• Platform coverage

  • TikTok reels feed
  • Instagram reels mode

• Instagram URL gating

  • Evaluation/blocker logic activates only on Instagram reels paths, not generic endpoints.

• Real-time decision execution

  • Reel-by-reel evaluation and actioning through IntersectionObserver.

• Full-screen blocking overlay

  • Black blocker while evaluating uncertain/skip flows.
  • Loader centered on the full viewport.

• Motivational quote panel

  • Random quote sourced from extension/assets/quotes.csv.
  • Displayed in right-third overlay panel.
  • Rotates every 5 seconds while blocker remains active.

• Skip reliability hardening

  • Multi-attempt "next" logic with pacing between automated scrolls.
  • Retry behavior to reduce "stuck blocker" edge cases.

• Runtime messaging resilience

  • Retry wrapper around extension messaging for MV3 wake-up races.
  • Special handling for "extension context invalidated" scenarios.
  • Graceful fallback path when messaging fails during evaluate flow.

• Watch tracking + event logging

  • Sends log_watch with:
    • action_type
    • watch duration
    • text payload
    • optional categories/deep analysis

• Action semantics

  • WAIT currently mirrors stay behavior (watch flow), but without auto-like click.

• Color credit system

  • Credits computed against recent watch events in Supabase.
  • Page-level grayscale intensity scales with remaining credits.
  • Credits consume for watched/stay-like actions (including WAIT and LIKE_AND_STAY in current logic).
  • Popup hint when credits are exhausted.

• Popup settings

  • Manage interests and blocked keywords.
  • Toggle autolike.
  • See current user id.
  • Edit user id from popup (validated + persisted).

• Parental lock integration

  • Popup enters read-only mode when lock state is active.
  • Lock state synced from Supabase on install/startup/alarm.

• FastAPI service

  • Endpoints:
    • POST /evaluate
    • POST /log_watch
    • GET /insights/{user_id}
    • GET /stats/{user_id}
    • GET /health

• Three-agent orchestration

  • Agent 1: Gatekeeper (fast action decision)
  • Agent 2: Deep analyzer (background semantic enrichment)
  • Agent 3: Insight generator (dashboard-level coaching synthesis)

• Gatekeeper policy prompting

  • Action-only output (score removed in current implementation).
  • Action rules tuned by interests and avoid keywords.

• Async deep-analysis cache + backfill

  • Deep analysis computed in background thread.
  • Event rows can be inserted immediately and patched later when analysis is ready.

• Supabase event pipeline

  • video_events storage
  • category/deep-analysis enrichment
  • per-user stats aggregation

• Health/status introspection

  • Agent/provider readiness and Supabase availability reporting.

• Main tabs

  • Neural Map
  • Analytics
  • AI Insights
  • Controls (Parental)

• User-scoped data view

  • Dashboard can load and persist user id for personalized analytics.

• Analytics + progress visuals

  • Skip/like/wait distribution
  • watchtime/quality metrics
  • rewiring progress indicators

• AI insights page

  • Narrative summary + recommendations generated from behavior data.

• Parental control management

  • Supabase-backed interests/blocked keyword editing
  • PIN lock/unlock flow
  • lock-state governed mutability
  • UI recently normalized to standard palette/text style

Create a .env in project root (example keys):

LAVA_GATEKEEPER_KEY=...
LAVA_ANALYZER_KEY=...
LAVA_INSIGHT_KEY=...
SUPABASE_URL=...
SUPABASE_KEY=...
USER_ID=...

Notes:

• USER_ID is not required for core backend operation (extension manages its own persisted user_id).
• Keep secrets out of committed files; rotate exposed keys.

From project root:

python main.py

API runs on http://127.0.0.1:8000.

cd dashboard
npm install
npm run dev

Dashboard runs on http://localhost:5173.

1. Open chrome://extensions
2. Enable Developer mode
3. Click Load unpacked
4. Select the extension/ directory
5. Reload the extension after code changes

FixMyFeed is designed to be strict where it matters (clear brainrot/toxic/avoid hits), flexible where it should be (uncertain-but-fine content), and reflective over time (insight dashboards + rewiring metrics).
Some flows are intentionally still evolving and may be refined as UX and model policy tuning continue.