PaperPilot

• 

  ui-dashboard.png

• 

  ui-webmcp-interaction.png

• 

  ui-latex-export.png

The Problem: Academic Research is Broken

Writing a literature review today requires researchers to download dozens of PDFs, manually copy-paste abstracts into a chatbox, and beg the model to summarize them. The result? The LLM hallucinates methodologies, loses track of citations, and breaks LaTeX math formatting.

I realized that LLMs don't need better prompts—they need ==Agent-Native infrastructure==.

Why this is a perfect fit for WebMCP

Before WebMCP, LLMs were blind to the user's UI. By implementing WebMCP, I exposed 13 strictly typed tools directly to the agent. This creates a massive UX paradigm shift:

• [x] Zero Copy-Pasting: The agent fetches full-text HTML directly via extract_findings.
• [x] No More Hallucinations: The compare_papers tool forces the agent to read structured matrix data.
• [x] Synchronous Collaboration: When the agent calls draft_section, the manuscript in my React UI updates instantaneously.

How I Built It: The Core Mechanisms

I built PaperPilot using Next.js 14, React 18, and TypeScript, deployed seamlessly on Vercel. To make this production-ready, I engineered three core mechanisms:

1. The Zero-Backend Event Bus

To achieve real-time UI synchronization without the latency of WebSockets, I implemented a decoupled event architecture. When ChatGPT executes a mutation tool: document.modelContext.registerTool({ name: "draft_section" }) It triggers a pure async Next.js Server Action. This action mutates the local data store and emits a native DOM CustomEvent (e.g., paperpilot:outlines-changed). The React UI listens to this bus and re-renders instantaneously.

2. Heuristic Extraction & Grammar-Aware Parsing

Academic HTML is notoriously messy. To prevent context window pollution, I engineered an extraction algorithm that isolates methodology while discarding boilerplate (Appendices, References). The extraction logic \(\mathcal{E}\) for a document \(D\) composed of sentences \(s_i\) is defined as:

$$ \mathcal{E}(D) = \bigcup_{i=1}^{N} s_i \quad \text{subject to} \quad \text{class}(s_i) \neq \text{"ltx_bibliography"} $$

Furthermore, naive sentence splitting ruins context by breaking on abbreviations like "e.g." or "et al." I parse sentences using a strict, grammar-aware Regex formula to guarantee abbreviation bypass: (?<=[a-z]{3,}[.!?])\s+(?=[A-Z0-9])

3. Math-Safe LaTeX Compilation

ChatGPT frequently generates inline mathematical formulas, such as \(\mathcal{O}(\log d)\). If I blindly escaped reserved LaTeX characters (like _ or %), it would break the math mode syntax. I engineered a custom AST-like parser that isolates \($\) and \($$\) blocks, safely escapes rogue characters only in standard text strings, and then restores the math blocks.

The agent also autonomously injects \cite{} markers linked to a dynamically generated \bibitem matrix, guaranteeing ==100% compilable Overleaf code==.

The Proof: My 20-Paper Stress Test

To prove PaperPilot isn't just a brittle hackathon demo, I built an automated stress test against 20 seminal AI papers (including Attention Is All You Need and Mistral 7B).

What I learned & What's next

I learned that strict typing in WebMCP inputSchemas is the absolute key to preventing agent hallucinations. Moving forward, I plan to expand my WebMCP toolset to integrate directly with Zotero and Mendeley, allowing researchers to bring their existing private libraries into the Agent-Native web.

Built With

• agents
• api
• arxiv
• automation
• chatgpt
• front-end
• git
• github
• javascript
• latex
• llm
• localstorage
• lucide-icons
• next.js
• openai
• parsing
• react
• regex
• research
• server-actions
• tailwind-css
• typescript
• vercel
• webmcp
• zero-backend