VEGA

Automated forensic auditing agent that filters web data using local vector embeddings before analyzing high-risk signals with Google Gemini. Core Architecture * Extraction: Hybrid scraping via Playwright (Dynamic/SPA) or BeautifulSoup (Static). * Filtration (Engineering Core): CPU-based vector search using FastEmbed (sentence-transformers/all-MiniLM-L6-v2) to score relevance against a "Risk Profile". Discards 90% of noise locally. * Analysis (Cognitive Nexus): Sends only high-scoring text chunks to Gemini 2.5/3.0 for structured risk assessment (JSON). Usage 1. Docker (Production) docker build -t vega-forensics . docker run -e ORT_LOGGING_LEVEL=3 vega-forensics 2. Python (Local) pip install -r requirements.txt python main.py Configuration (input_schema.json) | Parameter | Type | Description | |---|---|---| | startUrls | List | Target assets (HTML/PDF) to scan. | | risk_profile | String | Semantic target (e.g., "Undisclosed related party transactions"). | | scan_mode | String | dynamic (Headless Browser) or static (HTTP Request). | | gemini_api_key | String | Required for Phase 2 analysis. If omitted, runs in "Vector Scan" mode only. | Case Study: Enron Scandal Analysis Input: https://en.wikipedia.org/wiki/Enron_scandal Risk Profile: "Financial fraud and accounting irregularities" Output (Actual JSON): [{ "scan_data.risk_score": 98, "scan_data.status": "critical_risk", "url": "https://www.justice.gov/archives/opa/pr/bitconnect-founder-indicted-global-24-billion-cryptocurrency-scheme", "scan_data.summary": "Satish Kumbhani, founder of BitConnect, has been indicted for orchestrating a global $2.4 billion Ponzi scheme. He allegedly misled investors with a 'Lending Program' using a 'Trading Bot' and 'Volatility Software' that paid earlier investors with funds from later ones. Kumbhani also manipulated the price of BitConnect Coin and operated an unlicensed money transmitting business, evading FinCEN regulations. He is charged with conspiracy to commit wire fraud, wire fraud, conspiracy to commit commodity price manipulation, operation of an unlicensed money transmitting business, and conspiracy to commit international money laundering, and faces up to 70 years in prison. Kumbhani is currently at large.", "scan_data.next_step_query": "locate Satish Kumbhani or asset recovery for BitConnect victims", "vector_score": 0.47395562600420055, "scan_mode": "dynamic" }, { "scan_data.risk_score": 90, "scan_data.status": "critical_risk", "url": "https://www.sec.gov/enforcement-litigation/litigation-releases/lr-26369", "scan_data.summary": "The SEC concluded a civil enforcement action against Ripple Labs, Inc. and its executives, resolving violations of the registration provisions of the Securities Act of 1933. The action resulted in a final judgment imposing a $125,035,150 civil penalty against Ripple and an injunction prohibiting future violations, directly addressing an unregistered securities offering.", "scan_data.next_step_query": "details of Ripple Labs civil penalty and injunction terms", "vector_score": 0.4407136479739982, "scan_mode": "dynamic" }, { "scan_data.risk_score": 95, "scan_data.status": "critical_risk", "url": "https://www.pbs.org/wgbh/pages/frontline/madoff/etc/script.html", "scan_data.summary": "Avellino & Bienes were investigated and subsequently shut down by the SEC in 1992 for operating an unlicensed Ponzi scheme, with $441 million invested through Bernard Madoff. Despite the SEC having Madoff 'in their sights' and knowing he was operating as an unregistered investment advisor for 3,200 clients, they failed to take adverse action against him at that time, focusing instead on Avellino & Bienes.", "scan_data.next_step_query": "SEC investigation Bernard Madoff unregistered investment advisor 1992", "vector_score": 0.4148849254193477, "scan_mode": "dynamic" }] Scope & Limitations Evaluates contents Risk scores are analytical signals, not legal conclusions Outputs should always be reviewed by a human Intended Use Financial crime research Compliance / AML pre-screening Investigative journalism support OSINT triage If VEGA reduces 100 documents to 10 worth reading, it has done its job.