Netpost — LinkedIn Post Generator

A no-login LinkedIn post generator for B2B fintech and investment banking teams that writes only about what the industry actually published today — and shows its own working on every draft. Each click ingests ~75 live items across four sources, dedupes and ranks them, then returns five posts with their sources, a lint score and a similarity verdict attached.

▶ Live: netpost.rianfernando.com · How it works · llms.txt

Why it's different

Most AI writing tools start from a blank prompt and produce confident, sourceless copy. Netpost starts from the feed. It fetches live items from Hacker News, Reddit, industry RSS and News API on every click, fingerprints them to drop near-duplicates, and ranks what survives by keyword relevance × source weight plus a freshness bonus. Only the top 5–10 items become writing material.

Then it grades itself. Every draft is scored out of 100 by a linter that flags the specific tells of machine writing — weak hooks, vague claims, filler CTAs, missing credibility, unreadable paragraph length — and checked by a similarity engine against the style corpus and your previous posts. Anything scoring 45 or above is regenerated, not shipped. The score, the flags, the matches and the source links are all visible on the card.

Kept honest: there is no mock generation path. If the OpenAI or Supabase configuration is missing, the backend fails with a real error rather than serving placeholder content that pretends the system works.

Architecture

flowchart LR
  subgraph Sources["Live sources"]
    HN["Hacker News<br/>top 20"]
    RD["Reddit<br/>4 × 11 subreddits"]
    RS["RSS<br/>6 × 10 feeds"]
    NA["News API<br/>up to 20"]
  end
  subgraph Backend["FastAPI · Render"]
    IN["Ingest<br/>asyncio.gather"]
    DD["Dedupe<br/>SHA-1 fingerprints"]
    SC["Score<br/>relevance × weight + freshness"]
    GEN["Generate<br/>OpenAI structured JSON"]
    QC["Lint + similarity<br/>retry ≤ 3"]
  end
  subgraph Style["Style layer"]
    CO[("influencer_corpus.json<br/>pattern summaries")]
    PR["prompts/<br/>voice + generation"]
  end
  subgraph Store["Supabase · REST"]
    DB[("trends · batches<br/>posts · feedback")]
  end
  HN --> IN
  RD --> IN
  RS --> IN
  NA --> IN
  IN --> DD --> SC --> GEN --> QC --> UI["Next.js · Vercel"]
  CO --> GEN
  PR --> GEN
  SC --> DB
  QC --> DB
  DB -->|"feedback summary"| GEN

Loading

The frontend is a static Next.js App Router build; all live work happens in the FastAPI backend, which fans out to every source in parallel and persists to Supabase over its REST API (not the Python SDK, so modern sb_secret keys work cleanly).

What the app does

• Live trend ingestion — ~75 items per run from Hacker News, 11 finance/tech subreddits, 10 industry RSS feeds and News API, all fetched in parallel. Nothing is cached between clicks.
• Dedupe & scoring — SHA-1 fingerprints of normalized titles collapse syndicated stories; survivors are scored keyword relevance × source weight + freshness bonus (up to 2.0 for items under 24 hours old) and tagged Fintech / Automation / Banking.
• Trend brief — the top 5–10 items, each with a plain-language reason it ranked where it did.
• Five posts per batch — each mapped to a different trend, with hook, body, format, hashtags, tagging hints and source links.
• Two brand voices — founder (sharper, opinionated, operator-led) and company (measured, educational, category-authoritative).
• Anti-slop linting — nine flags scored out of 100: weak hooks, vague claims, generic filler, missing credibility, poor readability, hashtag spam, filler CTAs, corporate conclusions, excess tagging.
• Similarity checking — 35% token overlap + 40% 3-gram shingles + 25% cosine, with clear / review / blocked thresholds at 25 and 45.
• Anti-repetition — per-batch UUID nonce, fresh angle targets from eight banking-specific framings, recent-hook avoidance and trend-ID rotation, so clicking twice does not produce the same five posts.
• Performance feedback loop — record impressions, reactions, comments, reposts, saves and clicks; the summary by hook type, format and voice feeds the next batch.

The landing page

The home page renders a scroll-driven three.js scene of the pipeline itself: four source clusters stream items down a corridor, converge through a scoring core that dedupes and ranks, and resolve into five post panels whose stacked bars echo the Netpost mark. The camera flies a keyframed path as you scroll, with gentle bloom and pointer parallax. It is lazy-loaded, never server-rendered, drops to a single static frame under prefers-reduced-motion, and hides itself entirely if WebGL is unavailable — the page content stands on its own without it.

Data sources

Feed	Source	Auth	Weight
Top stories	Hacker News	none	1.1
Community	Reddit public JSON	none	0.95
Industry news	Finextra, PYMNTS, TechCrunch, American Banker, FT Banking, The Banker, Finovate, Tearsheet, Banking Dive, Payments Dive	none	1.2
Company blogs & changelogs	configurable via RSS_FEEDS_JSON	none	1.3
Headlines	News API	NEWS_API_KEY	1.4

Company blogs and changelogs are supported through configurable RSS entries with a source_type of company_blog or changelog, rather than a separate connector. Only headlines, summaries and links are stored — never full article text. See NOTICE.md for full attribution.

Run it

Backend

cd backend
cp .env.example .env          # add OPENAI_API_KEY, OPENAI_MODEL, SUPABASE_URL, SUPABASE_KEY
python3 -m pip install -r requirements.txt
uvicorn app.main:app --reload # http://localhost:8000
python -m pytest tests -q     # 7 tests

Frontend

cd frontend
npm install
cp .env.local.example .env.local   # NEXT_PUBLIC_BACKEND_URL, NEXT_PUBLIC_FEEDEX_KEY
npm run dev                        # http://localhost:3000
npm run lint                       # tsc --noEmit

GET /api/system/status reports whether AI and database configuration are actually ready — check it first if generation errors.

API

Method	Route	Purpose
GET	/health	liveness
GET	/api/system/status	AI + database readiness
GET	/api/style-guide	style bundle, voice guides, pattern summary
GET	/api/trends/brief	ranked trend brief with source breakdown
POST	/api/generate-batch	five posts for founder or company voice
POST	/api/feedback	record post performance metrics
DELETE	/api/trends/cleanup	prune stored trend events

Project structure

frontend/
  app/                   App Router pages, llms.txt, robots, sitemap, OG images
  components/            PipelineScene (three.js), Generator, PostCard, OgCard
  lib/site.ts            single source of truth for copy, FAQ and structured data
backend/
  app/api/               route handlers
  app/services/          trends, style, generation, linting, similarity, storage
  app/core/              settings and configuration
  tests/                 linting, similarity, generation parsing, storage
database/                Supabase schema + seeded style-pattern corpus
prompts/                 style guide, founder/company voice, generation prompt
docs/                    architecture notes, research sources, OG asset

Deployment

Layer	Host	Notes
Frontend	Vercel	root directory frontend, Next.js preset
Backend	Render	root directory backend, uvicorn app.main:app
Database	Supabase	private REST, server-to-server only

Canonical URLs are pinned to the subdomain via metadataBase, so any incidental *.vercel.app URL renders the same <link rel="canonical"> and search engines de-duplicate to netpost.rianfernando.com. Render's FRONTEND_URL is set to the same origin so CORS allows production traffic.

Discoverability

Route	What it serves
/llms.txt	machine-readable summary per the llms.txt convention
/robots.txt	names the major AI crawlers explicitly; allows /, disallows /api/
/sitemap.xml	both indexable pages
/opengraph-image	1200×630 PNG per page, generated by next/og

Structured data is split so nothing is declared twice: a sitewide WebSite + Person graph in the layout, WebApplication + FAQPage on the home page, and TechArticle on /how-it-works.

Feedback

In-app feedback is collected with Feedex, loaded from the root layout. The widget renders only when NEXT_PUBLIC_FEEDEX_KEY is set, so a checkout without the variable never boots it keyless. Theme is pinned to dark because the site has no light mode, and the accent matches the brand teal #5BC0BE.

The script uses strategy="afterInteractive" rather than the lazyOnload the Feedex docs suggest for Next. With lazyOnload the script is injected after the window load event has already fired — it still executes and fetches its remote config, but never attaches window.Feedex or renders a launcher, so no button appears.

Notes

• Supabase is required for real persistence and the feedback loop. The local fallback is off by default and only enables with an explicit ALLOW_LOCAL_DEV_FALLBACK=true.
• The seeded corpus in database/influencer_corpus.json stores pattern summaries only — no copied LinkedIn posts. Research references are cited in docs/research-sources.md.
• Generation requires a real OPENAI_API_KEY and OPENAI_MODEL. There is no mock path.

License

MIT © Rian Fernando. Independent project — not affiliated with or endorsed by LinkedIn Corporation. Netpost drafts posts and shows its scores; it does not verify claims. Read the sources before you publish.