Behavioral tracking
Batched, non-blocking event capture — page views, searches, dwell time, bookmarks. Uses sendBeacon on page hide.
SmartReco Build Challenge 2026
Your browsing behavior, turned into a personalized learning path.
SkillOrbit watches how you explore, search, and bookmark real courses — then uses AI to recommend your next step with a convincing story. Every suggestion is grounded in the live catalog, never hallucinated.
- 80+Catalog resources
- 7-stageLangGraph agent
- 100%Grounded retrieval
Live pipeline
Behavior → Qdrant → Mesh
1
Browse & searchEvents batched silently in background
2
Interest profile buildsCategories, dwell time, bookmarks weighted
3
AI path generatedQdrant retrieves · Mesh writes narrative
Grounded output
Build Production RAG Systems
Real catalog ID · validated before display
Product preview
See the full platform — every screen judges care about
Real screenshots from the live app: grounded recommendations, semantic search, LangGraph trace, and admin dual-write.
v-1-ora9.onrender.com/dashboard
How it works
Five simple steps from browsing to your next course
No magic black box — every step is visible, traceable, and backed by real data.
01
Explore the catalog
Search semantically, filter by career goal, open resources. Works without an account.
02
We watch quietly
Clicks, searches, dwell time, and bookmarks are batched and sent — never slowing you down.
03
Profile takes shape
Your interest radar updates — category weights shift as your behavior changes.
04
AI finds matches
Qdrant retrieves real catalog items. Mesh writes a personalized narrative from those results only.
05
Get your path
A clear next step, shareable link, full trace at /trace, and optional email delivery.
Browser
tracking.js batches events
Supabase
events + profiles + paths
Interest profile
category weights + triggers
LangGraph agent
7-stage pipeline
Qdrant
semantic retrieval
Mesh API
LLM + embeddings
Dashboard
path + trace + share
Resend
instant + weekly email
Features
Everything a real recommendation platform needs
Built for the SmartReco challenge — behavioral tracking, grounded AI, and production-grade design.
Semantic search
Qdrant-powered catalog search on /explore. Find resources by meaning, not just keywords.
LangGraph agent
7-stage pipeline: analyze → retrieve → evaluate → moderate → generate → validate → persist.
Grounded output
Every product ID validated against SQL before display. LLM writes narrative from retrieved candidates only.
Interest radar
Live category weights, streak tracking, weekly minutes vs goal — all driven by real activity.
Recommendation diff
Change your behavior, refresh your path, and see exactly what changed and why.
Shareable paths
Public /path/{id} links and PDF export — no login required to view shared results.
Full observability
/trace page shows pipeline stages, Qdrant scores, and Mesh trace IDs for every recommendation.
Admin dual-write
Add a resource in admin → automatically indexed in Qdrant. Sync health page detects drift.
Architecture
Production stack, not a prototype
Every layer is a real service — no mocks, no hardcoded recommendations.
Backend
FastAPI
Async Python API + Jinja2 server-rendered UI
Database
Supabase
PostgreSQL — auth, events, profiles, recommendations
Vector DB
Qdrant
Semantic search + RAG retrieval for the agent
LLM gateway
Mesh API
All chat + embeddings — GPT, Claude, Cohere via one key
Agent
LangGraph
Explicit multi-stage reasoning workflow
Email
Resend
Manual path emails + automated weekly digest
Scheduler
APScheduler
Daily digest check at 09:00 IST
Deploy
Render
Production hosting with health checks
LangGraph agent pipeline
analyzeBuild query from profile
retrieveQdrant vector search
evaluateScore threshold filter
moderateSafety check
generateMesh API narrative
validateCatalog ID check
persistStore + trace ID
Smart AI triggers: 5-minute cooldown · 24-hour TTL · regenerate only on meaningful behavior change
Bonus — proactive delivery
Recommendations that come to you
Two email flows powered by Resend — instant on demand, and automatic every 7 days.
Instant
Email me this path
One click on the dashboard sends your current recommendation — summary, next step, and grounded resources — straight to your inbox.
- Triggered by user on dashboard
- HTML email via Resend API
- Logged in
email_deliveries
Dashboard→Resend→Inbox
Automatic
Weekly learning digest
Every 7 days after onboarding, SkillOrbit checks if your path is stale. If behavior changed, it auto-generates a fresh recommendation and emails it — even if you never came back.
- APScheduler daily check at 09:00 IST
- External cron fallback for Render free tier
- Auto-generates path if missing or expired
- Duplicate guard via
delivery_kind
Scheduler→LangGraph→Resend→Inbox
Product preview
Dashboard built for explainable recommendations
Interest radar, live activity feed, streak tracking, and Mesh trace IDs — everything in one view.
skillorbit.app/dashboard
3Day streak
45/300Weekly minutes
12Signals
Live product UI — generate paths, share /path/{id}, inspect traces, and export PDF.
Career directions
Five learning orbits.
One path that fits you.
For judges
What to look for in 60 seconds
Every SmartReco requirement is implemented and visible — no hidden magic.
- ✓ Batched behavioral tracking — open DevTools → Network →
/api/events - ✓ Semantic search —
/explore→ search production RAG - ✓ LangGraph agent — generate path → open
/trace - ✓ Grounded output — every item has a real catalog ID
- ✓ Dual-write — admin add resource → appears in search
- ✓ Mesh API — trace ID in evidence panel → Mesh dashboard
- ✓ Weekly digest — APScheduler + Resend (see ARCHITECTURE.md)
- ✓ Guided demo — /demo → Auto-run demo
Ready when you are