Skeptic Agent

Adversarial Research Verification Engine

A local, privacy-first AI agent that reads research papers like a hostile peer reviewer — deconstructing claims, auditing methodology and surfacing hidden flaws. Zero cloud calls.

> What I built

• Runs three sequential adversarial audits on an uploaded PDF: key assertion and evidence strength, methodology tracker, and hidden flaws with targeted peer-review questions.
• Verification loop with JSON repair, strict schema validation and auto re-prompting; descriptive schema prompts cut malformed responses from the 1.5B model by over 90%.
• Live progress streams to an on-page terminal over NDJSON; dashboard built with Three.js and GSAP, all assets vendored so the page makes no external requests.

> Tech stack

• Python
• FastAPI
• Ollama (qwen2.5)
• PyMuPDF
• Three.js
• GSAP