Personalized AI & Technology IntelligenceLive

Astra

Personalized AI & technology intelligence, delivered Daily or Weekly.

Next.js 16React 19TypeScriptTailwind CSS 4shadcn/uiPrismaSupabasePostgreSQLRedisPythonOpenAIResendTurborepoCloud Run

Try live demo

Intelligence Pipeline

Astra

Interactive preview. Try the live demo for the full experience

3Cloud Run services

11Topic categories

Daily + WeeklyCadences

5Ranking signals

Overview

Astra ranks technology signal to your interests and delivers it as a personalized newsletter. The focus is the intelligence layer, not just filling the inbox. A Python collector continuously ingests articles from curated RSS, scrapers, and optional Exa sources, then enriches them with AI (summaries, topics, keywords, embeddings, quality scores) and deduplicates near-duplicates. A composite recommendation engine ranks content per user from topic affinity, engagement, freshness, diversity, and global quality, without relying solely on LLM prompts. Users get Astra Daily or Weekly digests via Resend, with open/click tracking feeding ranking.

Highlights

Python collector decoupled from Next.js, sharing a Prisma/Supabase PostgreSQL schema via SQLAlchemy

5-signal composite ranking (quality, topic affinity, engagement, freshness, diversity) with optional Redis cache

Pipeline: collect → enrich → embed → dedup → score → deliver (Daily/Weekly cadences)

Production on GCP Cloud Run (astra-web, astra-collect, astra-deliver) with Cloud Scheduler

Features

›SaaS landing with sample newsletter preview, how-it-works, and onboarding
›Magic-link authentication and weighted topic preferences across 11 categories
›RSS, scraper, and optional Exa collection with AI enrichment and heuristic fallbacks
›Embeddings-based near-duplicate detection plus URL/title fallbacks
›Personalized digests (Astra Daily / Astra Weekly) with HTML email and engagement tracking
›Admin console for sources, articles, worker runs, newsletters, and analytics
›Domain modules for auth, preferences, recommendation, newsletter, analytics, and unsubscribe
›Turborepo monorepo with shared scoring and branding primitives across TypeScript and Python
›Graceful degradation: Redis, OpenAI, Exa, and Resend optional for local development

Technical

Turborepo monorepo: `apps/web` (Next.js 16 App Router), `workers/collector` (Python feedparser + BeautifulSoup + OpenAI writing via SQLAlchemy to the Prisma schema), and packages for database, shared scoring/topics/branding, and shadcn/ui. Ranking blends five weighted signals in TypeScript; the collector enriches and embeds offline when OpenAI is unavailable. Production deploys three Cloud Run services with Cloud Scheduler driving collect and deliver.

Learnings

01Composite scoring beats prompt-only personalization for explainability, cost control, and tunable ranking
02Shared scoring primitives in `@workspace/shared` keep the web app and Python worker conceptually aligned
03Separating collect and deliver as Cloud Run jobs makes the intelligence pipeline operable without coupling to the web tier

All personal projects Open live demo