Blog

OpenAI Agents SDK: First Thoughts

code

agents

I spent last weekend exploring OpenAI’s newly released Agents SDK. To better understand its capabilities, I built a lightweight prototype illustrating how multiple LLM-based…

Eliciting Toxic Singlish from r1

red-teaming

responsible-ai

Cross-posted from GovTech AI Practice Blog

From Risk to Resilience: Adding LLM Guardrails From Day 1

guardrails

responsible-ai

Cross-posted from GovTech AI Practice Blog

Defence Against the Dark Arts: LLM Edition

llm

talk

responsible-ai

Here are the slides for a talk I gave at GovTech’s Data Science and AI Connect, a regular gathering of practitioners and enthusiasts across Singapore’s public service. Below…

Building Responsible AI — Why Guardrails Matter

guardrails

responsible-ai

Cross-posted from GovTech AI Practice Blog

Summarising Daily AI Papers with GitHub and Gemini

llm

talk

code

This blog post summarizes a talk I gave at Machine Learning Singapore. Thank you Martin and Sam for having me!

Generating Diverse Synthetic Data with GPT-4o

code

llm

Last weekend, I found myself with over $300 in expiring OpenAI API credits and decided to dive into generating synthetic data with GPT-4o. The result? A dataset of over 2…

A Simple Off-Topic Guardrail

code

In this post, we’ll explore how to build a straightforward off-topic detector using embeddings and K-Nearest Neighbors (KNN).

Navigating the limits of LLM

llm

talk

This blog post summarizes an online talk I gave for the Prompt Royale 2024.

Deploying HuggingFace models on AWS SageMaker with 2 commands

code

You can have a HuggingFace model up and running on SageMaker in just a few lines of code.

Fine-tuning TinyLlama with Axolotl and JarvisLab

llm

code

Fine-tuning TinyLlama to generate David Attenborough style narration - using Axolotl, JarvisLab, and OpenAI’s Batch API

Working with LLMs: A Data Analyst’s Guide

llm

talk

This blog post summarizes a talk I gave at GovTech’s Data Science and AI Connect, a regular gathering of practitioners and enthusiasts across Singapore’s public service.…

Side Quest at Build Together Hackathon: jina-clip-v1 and LanceDB Cloud

code

Jina AI released their new multimodal embedding model - jina-clip-v1 this week. It’s under an Apache 2.0 license and it can be downloaded/used via HuggingFace. You can find…

Notes for “Chameleon: Mixed-Modal Early-Fusion Foundation Models”

personal notes

llm

Here is the link to the original paper. These notes were prepared for my LLM Asia Paper Club sharing. Any feedback or areas for improvement would be most appreciated at cyzga…