Nasik Abbas

Evidence

33

Production releases shipped on the PadelGPT platform

169

Tools exposed to the agent tree across two MCP servers

1,280

Real support conversations in the evaluation corpus

67.8%

Share of support volume that needed an answer, not an action

The finding that reframed the project: the two largest blockers were self-service gaps in the product, not gaps in the AI.

65.1%

PadelGPT win rate against human club support

Limit: Measured but not validated — the LLM judge failed human calibration at Cohen's kappa -0.170, below chance, so this figure is reported for method and not as a result. The capability-gap analysis above does not depend on the judge.

565 of 703

Commits authored on the PadelGPT codebase