OMEGA — Real-Time Market Sentiment Intelligence

A full-stack platform combining NLP sentiment analysis, technical price prediction, and LLM-powered summarisation to give investors an information edge over raw news feeds.

OMEGA aggregates news articles for publicly traded companies, scores them across three sentiment axes (subjectivity, optimism, enthusiasm), generates LLM summaries, and surfaces everything alongside live price data and technical analysis signals — all in a single, real-time dashboard.

• Benchmarked five models (GPT-3.5, Llama2, Llama2-uncensored, Phi, Gemma) against a hand-labeled dataset of 30 articles using Spearman's rank correlation
• Selected Llama2-uncensored (ρ = 0.63) over GPT-3.5 for its open-source reproducibility and strong off-the-shelf alignment with human raters
• Characterized output variance per category across 100 repeated runs to quantify reliability before shipping

• Hybrid algorithm combining RSI, Parabolic SAR, and Fibonacci retracement levels (TA-Lib) to generate a directional confidence score (−4 → +4)
• Conservative signal design: returns 0 (uncertainty) in ~77% of cases; when confident, directional accuracy reached 100% across 7 live signals on AAPL
• Linear regression baseline (sklearn) achieving ~78% next-day accuracy using OHLCV features — with documented limitations around trend-following regimes

• LangChain refine chain for long-form document summarisation — handles articles exceeding model context windows by iterative chunk processing
• Temperature tuned to 0.2 to minimise hallucinations while preserving factual fidelity

• Articles scraped via BeautifulSoup at ingest time, not on request — eliminates per-request scraping latency
• Sentiment scores and summaries persisted to the database post-generation, so LLM inference runs once per article, not once per page load
• Average article download time: 0.12s; full 10-company, 6-article-each refresh under 7s worst case

• AbstractBaseUser with email as the unique identifier (overrides Django's default username model)
• JWT via simplejwt + React Context API with a checkLoginStatus server-round-trip to prevent client-side session spoofing

• Python 3.10+
• Node.js 18+
• Ollama with llama2-uncensored pulled (for AI features)

# 1. Backend
cd backend
python3 -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
python manage.py migrate
python manage.py runserver 8000

# 2. Frontend (new terminal)
cd frontend
npm install
npm run dev                      # → http://localhost:5173

The frontend proxies all /api/* requests to Django on port 8000.

AAPL · GOOGL · MSFT · NVDA · TSLA · AMZN · META · NFLX · JPM · BRK-B · DIS · PYPL

The system is structured as six decoupled Django apps (companies, users, articles, notifications, recommendations, auth), each with their own URL routing and views. This mirrors a microservices separation of concerns while remaining deployable as a monolith.

Rate limiting from Yahoo Finance (2,000 req/hr) is handled by spreading data refresh calls over time as background tasks. Transitioning to a distributed architecture (multiple Django workers + a dedicated data broker API) is a natural next step and is architecturally straightforward given the modular design.

The recommendation engine uses collaborative filtering: suggest companies followed by users with overlapping portfolios, with a sector-similarity fallback and a curated default list for cold-start users.

All backend functions have unit and integration test coverage. AI-generated outputs (getSentiment, summarizeText) were manually validated given their non-deterministic nature. The full test suite passes with 12/12 functions verified.

Backend → Railway: set the four env vars above and point Railway at backend/. The railway.toml and Procfile are pre-configured.

Frontend → Vercel: set VITE_API_URL to your backend URL. The vercel.json SPA rewrites are already in place.