Economics News Monitoring Agent

AI agent that tracks economics news for a target country, filters noise with Gemini, and returns structured, prioritized updates.

Stack

• Frontend/Backend: TanStack Start (React 19 + Vite, server functions)
• AI: Lovable AI Gateway → google/gemini-3-flash-preview
• Sources: Google News RSS, Bing News RSS, Yahoo News RSS, Reddit RSS
• Deployed on: Lovable (.lovable.app)

Core capabilities (per brief)

1. Monitor relevant sources — 4 RSS providers, fanned out in parallel across user-defined topics, sources, and competitors.
2. Filter noise — Gemini 3 Flash rewrites headlines, drops clickbait/dupes, scores importance.
3. Deliver clearly — structured cards grouped by importance, with category, why-it-matters summary, source, and link.
4. Configurable — country, topics, sources, competitors are all user inputs.

Key design decisions

Decision	Why
RSS over paid APIs	Zero cost, no keys, sufficient recall for a prototype
Multi-provider fan-out	Redundancy + broader coverage than any single feed
Title-normalized dedup	Cheap pre-LLM filter, saves tokens
Gemini 3 Flash	Cheap, fast, great at short structured extraction
Server functions (not edge)	Typed RPC, single deploy target, simple

Limitations

• No persistence — every scan is stateless
• No scheduling — scans are on-demand
• RSS feeds rate-limit and occasionally return empty
• No semantic dedup across near-duplicate stories
• Single-user, no auth
• Synchronous request (can be slow with many queries)

---

Further Development Roadmap

Week 1 — Reliability & data quality

• Add paid API fallback (NewsAPI or GNews) when RSS is empty/rate-limited
• Retry with exponential backoff per provider
• Semantic dedup with embeddings (cluster near-duplicates, keep best)
• Structured output via AI SDK Output.object for schema guarantees

Week 2 — Persistence & scheduling

• Lovable Cloud (Postgres) — store scans, updates, user configs
• Scheduled scans via pg_cron → hourly/daily digests
• Trend detection (item repeated across N sources = higher importance)
• Historical view: "what changed in the last 24h"

Week 3 — Delivery channels

• Email digest (Resend) — daily/weekly summary
• Slack/Telegram bot — push high-importance items in real time
• RSS feed of filtered output (meta!)
• Webhook for custom integrations

Week 4 — Multi-tenant & production polish

• Auth (Lovable Cloud) + per-user configs
• Rate limiting + usage quotas
• Admin dashboard: scan history, cost tracking, source health
• Evaluation harness: labeled dataset + precision/recall metrics on the filter

Later

• Sentiment scoring per update
• Entity extraction (companies, people, indicators)
• Chart overlays (correlate news with market data)
• Multi-country comparison dashboard
• Fine-tuned filter model on user feedback (thumbs up/down on updates)

Cost profile (current)

• RSS: free
• Gemini 3 Flash: ~$0.001–0.003 per scan (~80 items)
• Hosting: Lovable free tier

Run locally

```bash bun install bun dev ``` Requires GROQ_API_KEY in env (auto-provisioned on Lovable).