Subscribe

Responsible AI

A collection of 12 posts
Stress-Testing Government Communications with AI Personas Grounded in Real Voices

Stress-Testing Government Communications with AI Personas Grounded in Real Voices

Government communications can misfire in ways their authors never intended. We built an app that lets officers stress-test drafts against AI personas grounded in real Singapore voices, so they can catch what they missed in minutes rather than weeks.
EvalsResponsible AI
Building an automated Evals workflow that works (and open-sourcing it)
The Lab

Building an automated Evals workflow that works (and open-sourcing it)

How we built Kaleidoscope: A structured workflow for realistic, scalable, and human-aligned contextual AI evaluations.
Responsible AI
Yes, you’re absolutely right… Right? A mini survey on LLM sycophancy
The Lab

Yes, you’re absolutely right… Right? A mini survey on LLM sycophancy

Ever spoken to an AI and felt like it was responding with insincere praise?
Responsible AI
Benchmarking GPT-5 & GPT-OSS: A Responsible AI Approach
The Lab

Benchmarking GPT-5 & GPT-OSS: A Responsible AI Approach

Evaluating dimensions often overlooked by traditional benchmarks.
Responsible AIEvals

RabakBench: Multilingual AI Safety Evaluation Made Local

Global safety guardrails are often blind to local dialects and sensitivities.
Responsible AI

Does your LLM know when to say “I don’t know”?

Refusal by a model to answer may sometimes be more valuable.
Responsible AI

Fine-Tuning Language Models for Long-Context Data: Automated Stance Analysis of Citizen Discussions

Addressing technical challenges of processing high-volume public feedback for policy-making
Responsible AI

Securing Guardrails with Automated Red Teaming

Manual testing is no longer scalable.
Responsible AI
The Lab

(Part 2) LLM Safety Alignment for the Singapore Context using Supervised Fine-tuning and RLHF-based Methods

Safety must be "baked in".
Responsible AI
The Lab

(Part 1) LLM Safety Alignment for the Singapore Context using Supervised Fine-tuning and RLHF-based Methods

The process of "teaching" models to be safe
Responsible AI
The Lab

Eliciting Toxic Singlish from r1

A red-teaming exercise that proves even "reasoning" models can be coaxed.
Responsible AI