NOVA | Devpost

Setup wizard saves info to .env file (in gitignore) for local referencing
NOVA News, our mock site

Inspiration

Financial markets move fast, often faster than humans can react. News breaks across dozens of sites with mixed sentiment and vague headlines, leaving investors unsure how their portfolio will be affected.

We wondered "What if an AI system could continuously read financial news, understand it, and instantly tell you whether your stocks are at risk?"

That became the inspiration for NOVA, an automated, news engine. Entirely human written for human use.

What it does

NOVA is a tool that:

Scrapes live articles from financial news sites
Detects when a new article breaks
Reads the entire article (headline, metadata, text)
Uses a reasoning model to determine:
- Which portfolio stocks may be affected
- How positive or negative the article’s sentiment is
Automatically saves all findings in a structured JSON knowledge base
Tracks each writer’s historical accuracy (trust score) and sentiment tendencies
Works with a fully custom mock publication site (“Fahmy Macro Wire”) for live demo reliability

How we built it

Frontend / Mock News Website

Built by Fahmy using HTML and CSS, hosted on Vercel
Includes multiple demo articles with unique:
- Dates
- Editors
- Sentiments
- Tickers
Each article contains embedded JSON metadata for clean parsing and LLM development checking

Backend / Scraping System

Daniil built scrapers using Python + Selenium WebDriver
Extracts: headline, subheader, date, writers, and full article content
Currently scrapes CNBC (real world sample) and the custom Fahmy Macro Wire site (dev tool)

AI Reasoning Logic

Connor implemented the analysis pipeline using Google Gemini
For each article, the model outputs:
- Sentiment (–1 to +1 for most pessimistic to most optimistic)
- Likelihood of portfolio impact (0 to 1 for no impact to direct impact)
We introduced a scoring formula that approximates real financial influence:

$${Impact Score} = \text{sentiment} \times \text{trust}$$

Where:

sentiment = AI-estimated positivity/negativity of the article
trust = historical accuracy score of the writer or publisher

This formula allows NOVA to weigh article sentiment by how reliable the source historically is, instead of treating all news equally.

Onboarding Wizard

Runs when the project first opens
Prompts user for:
- Gemini API key (saved to .env)
- Portfolio tickers (saved as global config)
Ensures the system is fully ready with zero manual setup

Challenges we ran into

Financial news sites constantly change their CSS classes, breaking Selenium selectors
MarketWatch and similar sites have dynamic DOM structures with paywalls and overlays
Coordinating multiple scrapers required handling timing, waits, and failures
Ensuring the AI reasoning model always outputs strict JSON formatting
Designing a data structure that handles multi-source event grouping
Building a mock news site that looks real but is easy to parse for a live demo