Paul Nguyen

Paul Nguyen

Applied AI Engineer — Safety-Critical AI Products | Building Autodidact (local-first agent that learns) | PhD (Adversarial ML & Security)

Posts by Tags

ai agents

AI Agents: Buzzword or Paradigm Shift?

less than 1 minute read

Published:

“AI agent”—a fancy buzzword for AI model integration, or is there more to it? With increasing autonomy and decision-making capabilities, are we just rebranding models, or is this a true paradigm shift?

ai-agent

ai-safety

What Automotive AI Taught Me About Responsible Deployment

9 minute read

Published:

When you ship AI into a vehicle, the consequences of failure aren’t abstract. They’re not a bad user review or a dip in engagement metrics. They’re a driver who takes their eyes off the road because the system said something confidently wrong, or a voice assistant that crashes mid-navigation and leaves someone lost on an unfamiliar highway. Shipping AI where failures have physical consequences changes how you think about deployment — permanently.

applied-ai

Stop Training Your LLM Router: Zero-Shot Confidence Beats Supervised Baselines

5 minute read

Published:

TL;DR: We tested whether you actually need labeled training data to route queries between a cheap local LLM and an expensive cloud model. You don’t. Average token log-probability — available for free from the first query — matches supervised routing in-distribution and crushes it when the query distribution shifts. We tested across 3 model families, 2 datasets, ~4,500 queries, and $123 in cloud costs. All code and data are open.

artificial intelligence

AI Agents: Buzzword or Paradigm Shift?

less than 1 minute read

Published:

“AI agent”—a fancy buzzword for AI model integration, or is there more to it? With increasing autonomy and decision-making capabilities, are we just rebranding models, or is this a true paradigm shift?

autodidact

A transformer block from scratch: demystifying LLMs by building one

13 minute read

Published:

Many people see LLMs as a magic black box. At the architecture level, they’re simpler than they look. The core of every modern LLM - GPT, Claude, Gemini, Llama - is a transformer block. Stack one transformer block N times (typically 12 to 80), add an embedding layer at the input and an output head at the output, and you have a working ChatGPT-style LLM. That’s the whole recipe.

Stop Training Your LLM Router: Zero-Shot Confidence Beats Supervised Baselines

5 minute read

Published:

TL;DR: We tested whether you actually need labeled training data to route queries between a cheap local LLM and an expensive cloud model. You don’t. Average token log-probability — available for free from the first query — matches supervised routing in-distribution and crushes it when the query distribution shifts. We tested across 3 model families, 2 datasets, ~4,500 queries, and $123 in cloud costs. All code and data are open.

automotive

What Automotive AI Taught Me About Responsible Deployment

9 minute read

Published:

When you ship AI into a vehicle, the consequences of failure aren’t abstract. They’re not a bad user review or a dip in engagement metrics. They’re a driver who takes their eyes off the road because the system said something confidently wrong, or a voice assistant that crashes mid-navigation and leaves someone lost on an unfamiliar highway. Shipping AI where failures have physical consequences changes how you think about deployment — permanently.

awards

Best Student Paper Award at HOST 2020

less than 1 minute read

Published:

I’m excited to share that our paper “A Novel Golden-Chip-Free Clustering Technique Using Backscattering Side Channel for Hardware Trojan Detection” won the Best Student Paper Award at the 2020 IEEE International Symposium on Hardware Oriented Security and Trust (HOST).

calibration

Stop Training Your LLM Router: Zero-Shot Confidence Beats Supervised Baselines

5 minute read

Published:

TL;DR: We tested whether you actually need labeled training data to route queries between a cheap local LLM and an expensive cloud model. You don’t. Average token log-probability — available for free from the first query — matches supervised routing in-distribution and crushes it when the query distribution shifts. We tested across 3 model families, 2 datasets, ~4,500 queries, and $123 in cloud costs. All code and data are open.

cybersecurity

deployment

What Automotive AI Taught Me About Responsible Deployment

9 minute read

Published:

When you ship AI into a vehicle, the consequences of failure aren’t abstract. They’re not a bad user review or a dip in engagement metrics. They’re a driver who takes their eyes off the road because the system said something confidently wrong, or a voice assistant that crashes mid-navigation and leaves someone lost on an unfamiliar highway. Shipping AI where failures have physical consequences changes how you think about deployment — permanently.

georgia tech

hardware security

Best Student Paper Award at HOST 2020

less than 1 minute read

Published:

I’m excited to share that our paper “A Novel Golden-Chip-Free Clustering Technique Using Backscattering Side Channel for Hardware Trojan Detection” won the Best Student Paper Award at the 2020 IEEE International Symposium on Hardware Oriented Security and Trust (HOST).

hardware trojans

Best Student Paper Award at HOST 2020

less than 1 minute read

Published:

I’m excited to share that our paper “A Novel Golden-Chip-Free Clustering Technique Using Backscattering Side Channel for Hardware Trojan Detection” won the Best Student Paper Award at the 2020 IEEE International Symposium on Hardware Oriented Security and Trust (HOST).

learning-log

A transformer block from scratch: demystifying LLMs by building one

13 minute read

Published:

Many people see LLMs as a magic black box. At the architecture level, they’re simpler than they look. The core of every modern LLM - GPT, Claude, Gemini, Llama - is a transformer block. Stack one transformer block N times (typically 12 to 80), add an embedding layer at the input and an output head at the output, and you have a working ChatGPT-style LLM. That’s the whole recipe.

llm

A transformer block from scratch: demystifying LLMs by building one

13 minute read

Published:

Many people see LLMs as a magic black box. At the architecture level, they’re simpler than they look. The core of every modern LLM - GPT, Claude, Gemini, Llama - is a transformer block. Stack one transformer block N times (typically 12 to 80), add an embedding layer at the input and an output head at the output, and you have a working ChatGPT-style LLM. That’s the whole recipe.

Stop Training Your LLM Router: Zero-Shot Confidence Beats Supervised Baselines

5 minute read

Published:

TL;DR: We tested whether you actually need labeled training data to route queries between a cheap local LLM and an expensive cloud model. You don’t. Average token log-probability — available for free from the first query — matches supervised routing in-distribution and crushes it when the query distribution shifts. We tested across 3 model families, 2 datasets, ~4,500 queries, and $123 in cloud costs. All code and data are open.

local-llm

machine learning

AI Agents: Buzzword or Paradigm Shift?

less than 1 minute read

Published:

“AI agent”—a fancy buzzword for AI model integration, or is there more to it? With increasing autonomy and decision-making capabilities, are we just rebranding models, or is this a true paradigm shift?

open-source

research

Best Student Paper Award at HOST 2020

less than 1 minute read

Published:

I’m excited to share that our paper “A Novel Golden-Chip-Free Clustering Technique Using Backscattering Side Channel for Hardware Trojan Detection” won the Best Student Paper Award at the 2020 IEEE International Symposium on Hardware Oriented Security and Trust (HOST).

responsible-ai

What Automotive AI Taught Me About Responsible Deployment

9 minute read

Published:

When you ship AI into a vehicle, the consequences of failure aren’t abstract. They’re not a bad user review or a dip in engagement metrics. They’re a driver who takes their eyes off the road because the system said something confidently wrong, or a voice assistant that crashes mid-navigation and leaves someone lost on an unfamiliar highway. Shipping AI where failures have physical consequences changes how you think about deployment — permanently.

routing

Stop Training Your LLM Router: Zero-Shot Confidence Beats Supervised Baselines

5 minute read

Published:

TL;DR: We tested whether you actually need labeled training data to route queries between a cheap local LLM and an expensive cloud model. You don’t. Average token log-probability — available for free from the first query — matches supervised routing in-distribution and crushes it when the query distribution shifts. We tested across 3 model families, 2 datasets, ~4,500 queries, and $123 in cloud costs. All code and data are open.

talks

technology trends

AI Agents: Buzzword or Paradigm Shift?

less than 1 minute read

Published:

“AI agent”—a fancy buzzword for AI model integration, or is there more to it? With increasing autonomy and decision-making capabilities, are we just rebranding models, or is this a true paradigm shift?

transformer

A transformer block from scratch: demystifying LLMs by building one

13 minute read

Published:

Many people see LLMs as a magic black box. At the architecture level, they’re simpler than they look. The core of every modern LLM - GPT, Claude, Gemini, Llama - is a transformer block. Stack one transformer block N times (typically 12 to 80), add an embedding layer at the input and an output head at the output, and you have a working ChatGPT-style LLM. That’s the whole recipe.