Serenity Intelligence

Real-time AI investment intelligence. Serenity follows 17 investment and AI-infrastructure voices on X. It analyzes each post with LLMs: translation, summary, and bullish/bearish ticker extraction. Results go to a web dashboard and to subscribers on WeChat, typically within a minute of the original post.

It is a live product with paying users. I designed, built, and operate it end to end.

🔗 Live: alphaline.fit → click “Explore as guest” for a one-click, read-only demo (no sign-up).

At a glance

Sources tracked	17 X accounts, one channel each
Posts ingested and analyzed	9,000+ since May 2026 (live counter on the login page)
Tickers covered	1,300+
Customers	100+ paying subscribers to date
LLM robustness suite	152 / 152 cases passed across two rounds (reports)

Architecture

flowchart LR
    X[(X Filtered Stream)] -->|SSE| D[Stream daemon<br/>Railway · Docker]
    D -->|POST /api/stream/ingest| P[Ingest pipeline<br/>Next.js on Vercel]
    C[Cron · REST polling<br/>fallback mode] --> P
    P -->|dedupe by tweet id + SHA-256| DB[(PostgreSQL<br/>Supabase · Prisma)]
    P -->|Promise.any race| L1[DeepSeek]
    P -->|Promise.any race| L2[Xiaomi MiMo]
    P --> W[WxPusher → WeChat]
    DB --> UI[Dashboard · Alpha Map<br/>Supply chain · AI chat]
    F[Finnhub] --> UI

1. Ingest. A long-running daemon holds one connection to the X Filtered Stream and forwards each post to an authenticated ingest endpoint. A cron-driven REST poller with per-author adaptive intervals is kept as a fallback mode.
2. Analyze. Two LLM providers race in parallel. The first valid, schema-conforming answer wins.
3. Store and fan out. The analysis and ticker mentions are saved. Each channel's subscribers get a formatted WeChat message.
4. Read. A server-rendered dashboard shows the feed, a 30-day ticker leaderboard, a D3 force-directed "Alpha Map", a supply-chain view, and a streaming research chat.

Engineering highlights

Resilient streaming ingestion (scripts/stream-daemon.ts)

• Heartbeat watchdog: X sends a keep-alive every 20 s, and the daemon reconnects if nothing arrives for 40 s.
• Exponential backoff from 1 s up to 5 min. Stream rules re-sync from the database every 5 min.
• Fatal account errors, such as depleted API credits, stop the retry loop and alert the admin on WeChat instead of retrying forever.

Reliable structured output from LLMs (lib/ai.ts, lib/ai-normalize.ts)

• DeepSeek and Xiaomi MiMo run in parallel through Promise.any, so one provider's outage, quota, or malformed output doesn't block delivery.
• Model output passes through several layers: a state-machine sanitizer escapes raw newlines and curly or full-width quotes inside JSON strings, then jsonrepair, then Zod validation and normalization (confidence scaling, ticker dedupe).
• If every model fails, subscribers still get the original post rather than nothing.

Deduplication and push bookkeeping (lib/pipeline.ts)

• Posts are deduplicated by X id and by a SHA-256 hash of author + content, which catches re-posts and backfill overlap.
• Every push is recorded per (post, channel, payload hash), so retries and backfills skip messages that were already sent.

Auth, access control, and abuse protection

• Stateless sessions are HMAC-SHA256 signed and compared in constant time (lib/session-token.ts).
• Access is granted per channel with expiry dates, and synced from the WeChat subscription provider.
• Guest mode for reviewers: read-only, all channels unlocked, content delayed by 24 h, and no paid LLM calls. The delay cutoff is floored to the hour so guests share one cache entry.
• A sliding-window rate limiter on dashboard routes blocks scrapers (middleware.ts).

Frontend

• Next.js 15 App Router with React Server Components, and unstable_cache for the expensive aggregate queries.
• D3 force-directed graph with zoom and drag, a sticky sidebar layout, and a bilingual (EN / 中文) login page.

Screenshots

Alpha Map: ticker and narrative network	Ticker page: supply-chain position and earnings
Supply-chain heat map	Login with guest mode

Tech stack

App: Next.js 15, React 19, TypeScript, Tailwind CSS, D3 Data: PostgreSQL (Supabase, pgbouncer), Prisma AI: DeepSeek and Xiaomi MiMo (OpenAI-compatible APIs), Zod, jsonrepair. Anthropic and Gemini adapters are also included. Infra: Vercel (web and API), Railway (stream daemon, Docker), GitHub Actions (CI) Integrations: X API v2, WxPusher (WeChat), Finnhub

Project structure

app/                 Next.js routes: dashboard, auth, guest demo, API handlers
  api/stream/ingest  Authenticated endpoint the stream daemon posts to
  api/cron           Scheduled polling, retries, and maintenance
lib/                 Domain logic
  pipeline.ts        Ingest → dedupe → analyze → store → push
  ai.ts              Provider routing and the parallel model race
  ai-normalize.ts    JSON extraction, repair, and validation of model output
  channels.ts        Per-channel access control
  session-token.ts   Signed session tokens
prisma/              Schema and migrations
scripts/             Stream daemon and operational scripts (see scripts/README.md)
tests/               Vitest unit tests
docs/benchmarks/     LLM robustness test reports

Running locally

npm install --legacy-peer-deps
cp .env.example .env.local        # fill in DATABASE_URL, OPENAI_API_KEY, TWITTER_BEARER_TOKEN, ...
npm run prisma:migrate
npm run dev                       # http://localhost:3000

# Optional: run the real-time ingestion daemon
npx tsx --env-file=.env.local scripts/stream-daemon.ts

Quality checks

npm run lint
npm run typecheck
npm test                 # Vitest unit tests
npm run test:coverage

CI runs all of the above on every push and pull request.

Disclaimer

For information and research only. Not investment advice.

---

中文说明