← All speakers

Bio, Work & Ideas

Dimitrios Philliou

Conference affiliation: GitHub · 2024

Dimitrios Philliou is a senior growth product manager at Fivetran and former GitHub product manager who helped make AI coding assistants more extensible, interoperable, and secure. His work on GitHub Copilot Extensions and the GitHub MCP Server addresses a practical problem: how agents can access real software systems without compromising credentials, permissions, or human oversight.

Before GitHub, Philliou worked at Paystand and studied at Holberton School. He later worked on GitHub Actions before becoming a product manager for Copilot Extensions. In 2024, he joined a GitHub Copilot workshop at AI Engineer World’s Fair and helped developers build custom Copilot extensions.

His subsequent work on GitHub’s open-source MCP server connected AI clients to repositories, issues, and pull requests, with particular attention to three product decisions:

  • Risk-aware agent permissions: Tool annotations distinguish read-only operations from destructive actions, allowing routine inspection while requiring approval for consequential changes.
  • Credential protection against prompt injection: Secret scanning and push protection inspect tool-call inputs before execution and block requests that could expose credentials.
  • Governed interoperability: The remote GitHub MCP service combines compatibility with multiple AI clients, OAuth 2.1 with PKCE, and centralized organizational policies.

In May 2025, Philliou also co-created Cotext.ai, a hackathon prototype exploring scoped-access personal agents connected to external services through MCP integrations.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

A hands-on path through Copilot in VS Code, from supplying useful context and reviewing JavaScript suggestions to understanding model latency, extensions, and multistep development.

  • What does Copilot need to help you write code?
    0:30 ↗
  • From editor context to a suggestion
    3:07 ↗
  • Make the request—and its context—explicit
    7:21 ↗
  • Good names and concrete examples reduce ambiguity
    14:22 ↗
  • Prepare the workshop environment
    19:27 ↗
  • Find the right participant and open a JavaScript file
    29:33 ↗
  • Ghost text, inline edits, and conversation
    36:22 ↗
  • Accept only the part you want
    41:22 ↗
  • Generate a URL parser, then inspect the repair
    47:00 ↗
  • Explain code, reset context, and generate tests
    50:36 ↗
  • Model quality has to fit the interaction
    59:30 ↗
  • Extend chat without assuming full model replacement
    1:04:22 ↗
  • Updated knowledge and automatic delegation solve different problems
    1:07:24 ↗
  • From a reviewed suggestion to a reviewed plan
    1:13:33 ↗

References