ni5arga/deanonymizer

Deanonymizer is a command-line tool that estimates how identifiable a person is based on their publicly visible activity on platforms like Reddit, Hacker News, GitHub, and Stack Overflow. The stated purpose is defensive: it is designed to help people measure their own exposure before a real adversary does it for them. The tool does not use private APIs or credentials, only information that is publicly accessible. The way it works is to collect posts, comments, profile fields, and metadata for a given username across whichever platforms the user specifies. It then runs two types of analysis in parallel. One is an AI model pass that looks for patterns suggesting real-world identity, such as mentions of a city or employer, references to a personal routine, disclosed personal details, or links to other accounts. The other is a rule-based pattern matcher that looks for email addresses (including obfuscated versions like name-at-domain) and usernames from a list of over a dozen platforms. Results from both passes are combined into a ranked risk report. The report groups findings into high, medium, and low confidence categories, with links back to the specific posts or comments that produced each finding. Concrete items like emails and discovered handles are surfaced in a dedicated section at the top, separate from the AI-generated inferences. The report can be saved as JSON for tracking changes over time. There is also a strict mode that fails the audit if no externally verifiable evidence exists for a claim. For the AI analysis, the tool supports several backends. By default it uses Anthropic's Claude API, but it can also route through any OpenAI-compatible endpoint, including local models via Ollama, Google's Gemini, or Groq. There is also a mode that works through the Claude Code command-line tool, which requires no API key if the user already has an active Claude Code session. The README notes the academic research that informed the design, specifically the observation that individual disclosures that seem harmless in isolation can become identifying when combined across many posts or platforms.