Mobius-Screen — AI equities screener on Nebius Serverless AI

An open, reproducible stock screener that ranks an F&O / equities universe by combining macro-thematic context + Elliott-Wave structure + classical technicals and asks an LLM to produce ranked, risk/reward-gated trade ideas as strict JSON. Inference runs on Nebius Serverless AI — a hosted/dedicated Endpoint for live screens, and a containerized Serverless AI Job for batch ingest + screening.

Built for the Nebius Serverless AI Builders Challenge. Blog post tag: #NebiusServerlessChallenge.

Why Serverless AI fits this problem

Screening is bursty and embarrassingly batchable. A market is interesting for a few hours a day; the rest of the time you want zero idle GPU cost. Two Nebius products map cleanly onto the two access patterns:

• Serverless AI Endpoints — an OpenAI-compatible LLM endpoint backs the live "Screen now" button in the UI. Pay for inference, not idle capacity.
• Serverless AI Jobs — a one-off/scheduled container refreshes the price cache for the whole universe and runs the screener end-to-end, with per-item retry and per-market isolation. Exactly the batch shape Jobs are built for.

Architecture

  Browser ──▶ Next.js UI (Cloudflare or Docker)
                 │  POST /api/screen
                 ▼
        src/lib/llm.ts  ──selects provider──▶  Nebius Serverless AI Endpoint
                 ▲                              (OpenAI-compatible /v1, Qwen2.5)
                 │                                       ▲
   Nebius Serverless AI JOB ───────────────────────────┘  (batch screens)
   (jobs/batch-screen.ts)
                 │  ingest prices + indicators
                 ▼
            Turso (libSQL)  ◀── reads ── Next.js UI

• Provider: src/lib/llm.ts talks to Nebius via the nebius OpenAI-compatible ChatProvider. Because it's configured purely by NEBIUS_BASE_URL + NEBIUS_API_KEY, you can point it at a dedicated Serverless AI Endpoint, Token Factory, or any OpenAI-compatible server with no code change. No proprietary AI SDK is used.
• Data: yahoo-finance2 (NSE/DFM) + TradingView scanner (ADX); RSI/MACD/SMA via technicalindicators, Fibonacci computed manually.

Quick start (local)

npm install
cp .env.example .env.local      # fill in NEBIUS_API_KEY, TURSO_*, AUTH_*
npm run dev                     # http://localhost:3000

The screener needs a populated price cache. Either run a screen (it tops up missing symbols live) or run the batch job once: npm run job.

Using Nebius Serverless AI

Option A — Token Factory (hosted, recommended for cost/repro)

No GPU to manage; pay per token. Get a key from the Nebius console, then set:

NEBIUS_BASE_URL=https://api.tokenfactory.nebius.com/v1
NEBIUS_API_KEY=...          # your Nebius key
NEBIUS_MODEL=Qwen/Qwen3-30B-A3B-Instruct-2507

Option B — dedicated Serverless AI Endpoint (vLLM)

Deploy your own model endpoint and point the app at it:

bash scripts/nebius/deploy-endpoint.sh        # prints endpoint IP + AUTH_TOKEN
# then:
NEBIUS_BASE_URL=http://<endpoint_ip>/v1
NEBIUS_API_KEY=<AUTH_TOKEN>
NEBIUS_MODEL=Qwen/Qwen2.5-7B-Instruct         # size to your GPU preset

Cost note: large dense models need multiple GPUs on a self-hosted endpoint. Use a small/MoE model for Option B, or Token Factory (per-token, no GPU to manage) for the default Qwen3-30B-A3B-Instruct at very low cost.

Run the batch Serverless AI Job

# cheap local test run — 3 symbols, one market (uses your .env.local)
MARKETS=UAE MAX_SYMBOLS=3 npm run job
# full run
npm run job

# on Nebius (build → push → create job)
REGISTRY=cr.<region>.nebius.cloud/<project> SECRET_ID=<mysterybox-secret> \
  bash scripts/nebius/run-job.sh
nebius ai job logs <job_id> --follow

Run with Docker

docker build -t mobius-screen .
docker run --rm -p 3000:3000 --env-file .env.local mobius-screen

Configuration

Variable	Required	Default	Notes
NEBIUS_BASE_URL	yes	Token Factory /v1	or http://<endpoint_ip>/v1
NEBIUS_API_KEY	yes	—	Nebius key or endpoint token
NEBIUS_MODEL	no	Qwen/Qwen3-30B-A3B-Instruct-2507	any served model
TURSO_DATABASE_URL / TURSO_AUTH_TOKEN	yes	—	libSQL DB
AUTH_SECRET, AUTH_GOOGLE_ID, AUTH_GOOGLE_SECRET, AUTH_URL	yes	—	Google sign-in
MARKETS	job	IN UAE INTL	markets the job screens
MAX_SYMBOLS	job	0	cap symbols/market (0=all); set small for a cheap test
BATCH_USER_ID	job	batch	owner of job-written screens

Full deploy details: docs/DEPLOYMENT.md.

Hardware & runtime

• App / Job container: CPU only (data fetch + LLM call are HTTP). A full-universe job run is dominated by upstream data latency + one LLM call per market.
• Endpoint (Option B): GPU (e.g. gpu-l40s-a / 1gpu-8vcpu-32gb), sized to the model.
• Approx cost: Token Factory inference is a few cents per screen at 72B; a self-hosted endpoint bills GPU-time while running. Record your measured numbers in your blog post (judges expect a cost figure).

Expected output

POST /api/screen (and each job market) returns ranked picks:

{
  "generatedAt": "2026-06-08T09:30:00.000Z",
  "macroNote": "Rate-cut tailwind for rate-sensitives; metals cooling.",
  "topBuys": [
    {
      "symbol": "RELIANCE", "side": "BUY", "conviction": 4,
      "thesis": "Energy/retail theme intact; Wave-4 ABC complete.",
      "waveCount": "Wave 4 ABC complete, Wave 5 starting",
      "entry": 1450, "stop": 1390, "target": 1620,
      "timeframe": "1-3 months", "risks": ["Loses 1390 support"]
    }
  ],
  "topSells": []
}

Project map

Path	What
src/lib/llm.ts	provider dispatch + shared screen orchestration
src/lib/providers/	nebius OpenAI-compatible provider
src/lib/prompts.ts	system/user prompts + JSON schema
src/lib/markets.ts	market registry (add markets/stocks here)
jobs/batch-screen.ts	Serverless AI Job entrypoint
jobs/Dockerfile · Dockerfile	job image · app image
scripts/nebius/	endpoint deploy + job run scripts

License

MIT. Not investment advice — outputs are research candidates, not signals.