Built around the practice of doing the work, not describing it
Backprop Lab was set up in Kuala Lumpur by engineers who found that most available AI education either stopped too early or started from the wrong place. The courses here are designed to fix that.
Back to HomeWhere Backprop Lab came from
Backprop Lab started with a specific frustration. The people who wanted to move from data analysis or backend work into model development could find plenty of overview material, but almost nothing that required them to write code that actually ran, failed, and had to be debugged in the way production code does.
The first course — Python and Data Engineering for AI Work — grew out of a workshop series run for a small group of engineers in the KL technology community in late 2023. The workshop was ten weeks, required six to eight hours of work per week, and included written code review on every assignment. The feedback was consistent: the format worked because it did not let you move on until the previous cell produced the right output.
The Transformer Architectures course followed in 2024, built for the engineers who had completed the first course and wanted to implement attention mechanisms from the paper rather than call a library. That course now includes shared cluster access, sixteen assignments, and weekly paper-reading sessions. It takes eighteen weeks and about twelve to fifteen hours per week.
The AI Systems Engineering Capstone Track is the most recent addition. It is aimed at practitioners who already know how to train models and need to build the surrounding system — retrieval, serving, evaluation, monitoring, and cost control — to a standard that holds up under real load. It runs for twenty-six weeks with a dedicated technical mentor and fortnightly architecture reviews.
The school operates from Suite 8-3, Menara Suezcap, KL Gateway. The team is small. Everyone involved in course delivery has written and shipped production AI systems.
The team
Ahmad Farouk
Seven years building data pipelines at scale across logistics and fintech in Malaysia. Teaches the Python and Data Engineering course and reviews all first-course assignments.
Siti Liana
Worked on language model pretraining and fine-tuning for three years at a regional NLP research group before moving into education. Leads the Transformer Architectures course.
Rajan Kumar
Systems engineer with a background in inference infrastructure and deployment. Runs architecture reviews on the Capstone Track and co-authored the evaluation methodology module.
How we maintain course quality
Written assessment record
Every learner receives a written record of completed assignments and their written code reviews. The record is produced at course close and does not expire.
Code review on every submission
Assignments are not auto-graded. Each submission receives written feedback addressing correctness, readability, and engineering choices from an instructor who has written production code in the relevant area.
Personal Data Protection Act compliance
Learner data is handled in compliance with Malaysia's Personal Data Protection Act 2010. We do not share enrolment or personal data with third parties for commercial purposes.
Source-level curriculum
Course material references primary sources — papers, specifications, and documentation — rather than secondary summaries. Where a concept comes from a paper, learners read the relevant section of that paper.
Known failure modes documented
Each course unit lists where learners commonly get stuck, what the failure looks like in the output, and how to work through it. This list is updated when new patterns appear across cohorts.
External critique on Capstone work
The Capstone Track's evaluation methodology is subject to external critique from a second reviewer outside the instruction team. The defensive review exercise is conducted by someone other than the primary mentor.
What we focus on and why
AI development education in Malaysia has expanded considerably over the past three years, but most of it concentrates on awareness and tooling rather than on the underlying engineering. Backprop Lab works at the engineering level: the code that processes data, the mathematical machinery of attention, the infrastructure that serves a model and keeps it observable in production.
This means the courses ask more of learners in terms of time and prior knowledge than an overview programme would. The Python and Data Engineering course expects comfort with Python. The Transformer Architectures course expects PyTorch and calculus. The Capstone Track expects experience training models. These are real prerequisites, not suggestions.
The advantage of working at this level is that what you learn is not tied to a specific library version or API. Attention as a mechanism does not change when a framework updates. Memory profiling as a discipline applies whether you are working with Pandas or Polars. The engineering principles transfer.
Backprop Lab also takes the question of responsible deployment seriously as an engineering matter. Dataset licensing, consent, provenance, and model risk are addressed in the curriculum not as ethics modules separate from the technical content, but as constraints that affect how you design and evaluate a system. The Capstone Track includes a module on model risk and a written incident post-mortem because these are part of what it means to build a system that works.
Read through the course details and then get in touch
The enquiry form on the main page is the right place to start. We will respond with a plain answer to your questions about prerequisites and fit.
Send an Enquiry