Kevin Tan
Articles sur Kevin Tan
PDF Prompt Injection: Hidden Text Only Your AI Agent Sees
Hidden white-text prompts in PDFs fooled AI peer reviewers. How a deterministic geometry scan detects invisible instructions before your agent reads them.
MCP File Upload: Why Base64 Tool Arguments Can Corrupt Files
An MCP file upload sent as a base64 tool argument is retyped by the model, and a 3 KB PNG arrived broken with a success response. Here is the fix that held.
Extract Chart Data From PDFs: An MCP Tool That Won’t Guess
An MCP tool that reads exact chart values straight from a PDF's vector geometry, then hands the ambiguous judgment calls back to the LLM instead of guessing.
How to Run LLM-as-a-Judge Cheaply Without Biasing Results
How to cut LLM-as-a-judge calls and tokens from the CLI: sequential ballots, stripped context, and measured run-to-run instability to catch biased shortcuts.
How Benchmarking Against Amazon Bedrock Improved My RAG
I benchmarked my RAG against stock Amazon Bedrock Knowledge Bases and lost. The failures exposed two bugs I never tested for, and fixing them made it better.