Khmer TTS
Adapting an English text-to-speech model to Khmer on a small budget: new tokenizer, number reading, and evaluation with a native listener in the loop.
View projectAI / Machine Learning Engineer
Richard Cheam
Currently at Renault. Engineer's and Master's degree in Data Science & AI.
Selected Projects
DecisionOps
Offline evaluation of learned and fixed policies in a bounded diagnostic workflow, with replayable traces and a standalone report viewer.
View projectInference Engineering
Public field guide to LLM inference engineering: ten question-led chapters and 16 interactive figures computed from tested functions over pinned sources.
View projectEmbedding Diagnostics
Controlled study of which label-free diagnostics detect a degenerated self-supervised embedding, on CIFAR-10 and BDD100K driving scenarios.
View projectEnigmatica
Chapter-based cipher puzzle game inspired by manga decoding challenges, built with reusable Python mechanics and a public GitHub Pages interface.
View projectXAI Vision Transfer
Checkpoint-first computer-vision study on explainability, transfer learning, and architecture comparison across CNN, ViT, and DHVT.
View projectCustomer Feedback Intelligence
Rebuilt an IMDb sentiment project into a public customer-feedback dashboard with benchmarked inference, transfer checks, triage, and exportable batch analysis.
View projectSSL and Semi-supervised Learning on MNIST
Reached strong accuracy under extreme label scarcity with pseudo-labeling and SimCLR representation learning.
View projectBlog
Technical blog posts and explainers
Tokenization: what a tokenizer decides before training starts
The tokenizer is trained before the model and fixed after it, and it quietly decides how much every sentence costs. Follow...
Read postSupervised fine-tuning: teaching a base model to answer
A pre-trained model continues text; it does not answer. Supervised fine-tuning teaches it to, with the same next-token los...
Read postA complete small training run: from raw text to a checkpoint you can trust
Train a one-million-parameter language model on a laptop CPU, end to end: tokenizer, data splits, model, AdamW with warmup...
Read postSharding and recomputation: making a model fit
Data parallelism gives every GPU a full copy of the training state, which stops working once that state outgrows a GPU. Le...
Read postPaper Reviews
Research takeaways worth revisiting
Chinchilla: how much data a model should see, and how firmly the paper shows it
Chinchilla argues that parameters and training tokens should grow together with compute, roughly 20 tokens per parameter at its optimum...
Read reviewDeepSeekMath and GRPO: a cheaper RL algorithm, and what RL actually changed
DeepSeekMath introduced GRPO, which replaces PPO's value model with the average reward of a group of samples. The paper's lasting value...
Read reviewDPO: preference learning without a reward model, and how strong the evidence was
DPO shows that the optimal policy of the KL-regularized reward objective defines its own reward, so preferences can train the policy di...
Read reviewContact