Vale construir sua versão de Bazzly?

Build a local Reddit opportunity monitor called reddit-radar. No hosting, no accounts, no telemetry.

Stack, no substitutions:
- TypeScript on Node 20, run with tsx
- SQLite via better-sqlite3 for storage
- snoowrap for the Reddit API
- A single Hono server on port 8787 serving a plain server-rendered HTML review page, no React, no build step
- Secrets in .env: REDDIT_CLIENT_ID, REDDIT_CLIENT_SECRET, REDDIT_REFRESH_TOKEN, REDDIT_USER_AGENT, OPENAI_API_KEY

Config lives in config.json at repo root: product name, one paragraph product description, list of subreddits to watch, list of search keywords, and a minimum relevance score from 0 to 100.

What it does:
1. A poll command fetches new posts from each watched subreddit and results for each keyword from Reddit search, deduping by post id against SQLite.
2. For each new post, one LLM call returns strict JSON: relevance score 0-100, a one line reason, whether the post is actually asking for a recommendation, and a suggested reply of at most 120 words written as a normal human who happens to have built the thing.
3. Store everything. Anything at or above the minimum score lands in a review queue.
4. The web page lists the queue newest first: subreddit, title, score, reason, link to the thread, the draft reply in an editable textarea, and buttons for Copy, Dismiss, and Mark handled.

Explicitly out of scope, do not build these: posting or commenting through the API, DM sending, multiple account support, proxy rotation, scheduling of posts. This tool suggests, the human posts. Print a warning at startup saying exactly that.

Also include: respect Reddit rate limits with a simple queue and backoff, an npm script for a cron style loop every 30 minutes, a seeded demo mode that works with fake data and no API keys, and a README with how to get a Reddit refresh token for a script app.

When done, run the demo mode and show me the review page output.