"Sahih - Arabic quality gate on ROCm team on AMD AI League Hackathon"
Team Idea
An agentic quality gate for Arabic in AI pipelines, running open models on AMD GPUs with ROCm. It audits a pipeline end to end — corpus, tokenizer, generation, rendering, text extraction — and reports every point where Arabic silently degrades, with reproducible evidence. The agent plans its own audit, runs probes against real models, then verifies each finding against text it did not generate before reporting it. Built on measured prior work: 341 public Arabic datasets audited and published under CC BY 4.0 with a DOI, and upstream fixes merged into pypdf, matplotlib, Pillow and sentencepiece. Arabic failures in AI systems are dangerous because they look fine to anyone who does not read Arabic, so they pass review and ship.
Team Members
.png&w=256&q=75)
syamjith NK
Syamjith-NK
Submission
No submission yet