Mohammad Khursani bin Mohd Shaib

I make machines understand
how Malaysians actually speak_

Research engineer, speech models and machine learning. Six years of ASR, benchmarks, LLM fine-tuning and the failure analysis nobody else shows you.

The flagship: Aisyah, and the benchmark that kept it honest

At Revolab I lead the speech team. Aisyah 1.0 Pro reached 5.79% word error rate on our combined 1,899-sample Malaysian benchmark: 9.55% on the telephony subset, the hard one. Compressed calls, short utterances, bad microphones: the places real speech goes to die.

The benchmark behind it spans 12 domains: 820 public samples, 1,079 held-out, with second-pass review of every flagged sample and spoken-form references, so formatting differences never masquerade as recognition errors.

pergi ke kedai selepas hujan redarela petang tadi illustrative: the word-level diff view from our evaluation explorer: the model heard rela, the reference says reda

benchmark design → 14-model leaderboard → how we built it → source →

Running right now

Things I build tend to stay running. This page checks.

Semak community-verified incident board Revolab Speech API Malay ASR + TTS in production Malaysian LoRA collection public adapters on Hugging Face

Experience

  1. Jun 2025 to now
    Lead AI Researcher, Revolab

    Multilingual TTS and ASR across Malay, Chinese, English and Levantine Arabic, extending to Tamil and Indonesian. Automated labelling from ingestion to gold-layer datasets. TTS latency and time-to-first-audio. End-to-end speech-to-speech research. Five researchers and engineers through model-development experiments.

  2. Jun 2023 to Jun 2025
    Associate Manager, Amanah Saham Nasional Berhad

    Fine-tuned LLMs on company knowledge without retrieval, plus a RAG chatbot from official content. MLflow pipeline for training and experiment tracking. Text classification, warehouse pipelines, PII-restricting database views.

  3. Mar 2020 to Jun 2023
    Senior Developer · AI & ML Engineer, Omesti

    Bahasa Malaysia speech recognition and audio-acquisition pipelines; active learning guided the cleaning. W&B hyperparameter sweeps on 10% of data, full-data training on the winners. MLOps for a recommender with explainability and confidence intervals. Document object detection; number-plate recognition.

  4. Sep 2017 to Feb 2020
    Full Stack Software Engineer, Vase.ai

    Backend and dashboards for mobile and web products; Node.js algorithms; a PowerPoint-from-dashboard service on AWS Lambda.

  5. May 2016 to Dec 2016
    Software Engineering Intern, IPG Mediabrands

    C#/Selenium scrapers, Docker containers on AWS ECS, automated UI and API tests.

BTech (Hons), Information and Communication Technology, Universiti Teknologi PETRONAS. Languages: Malay, English. Kaggle Competitions Expert.

Public artifacts

Hugging Face: Malaysian LoRA collection → NeMo: QuartzNet checkpoint migration and NaN-loss fix → the benchmark source code → Kaggle → GitHub →

Say hello

khursani8@gmail.com LinkedIn → +60 18-315 4932

Puchong, Selangor, Malaysia.