Artificial Intelligence ๐Ÿค–

OpenAI API
NLP
LLMs
AI in Games
AI in Education
โ€œ
Robots have solved and will continue to solve so many human problems. Except for all the ones that they cause.
Hank Green

I have practical hands on experience folding A.I. features into applications.

I've used the OpenAI API as a reasoning engine for a Language Learning App at Arti.

I've employed Azure AI Speech to build out a voice controlled vocabulary builder.

Skills Showcase

๐Ÿ“ฑ

App Development

๐Ÿ“ฑ

App Development

โ€œ
If we want users to like our software, we should design it to behave like a likeable person.
Alan Cooper

Many of my most recent projects have been at the intersection of games, education and apps.

  • Built gamified app for language learners: Arti Languages
  • Utilized the OpenAI API & Azure Natural Language Processing to build voice-controlled vocab builder for language learning.
  • Currently building Field Guide, a 'nature appreciation app' that augments open source datasets to teach about the natural world in fun ways.
๐Ÿฆพ

Software Development with AI

๐Ÿฆพ

Software Development with AI

โ€œ
Artificial intelligence is not a substitute for human intelligence; it is a tool to amplify human creativity and ingenuity.
Fei-Fei Li

I hold certifications in Generative AI for Software Development and Generative AI with Large Language Models, I utilize LLMs effectively across documentation-writing, refactoring, pattern selection and debugging. That said I use it more sparingly when writing code, because it's fundamentally important that I understand the code in use to stop technical debt down the line.

โ€

๐Ÿ’ฌ

Azure AI Speech Services

๐Ÿ’ฌ

Azure AI Speech Services

โ€œ
Human languages tend to be much more ambiguous than computer languages because humans are much smarter about interpreting the context.
Larry Wall

When building the Voice-controlled vocabulary builder at Artiย  I dug into the weeds of Azure AI Speech.

Speech synthesis in multiple languages holds unique challenges, for example so called multi-lingual models often don't pronounce non-English words correctly, especially if you switch languages half-way through a sentence.

So to take a simple example, if you want to synthesise "What is the meaning of un sous sol?" - which means basement. One can't rely on a multi-lingual model to identify the 'code switch' instead the text needs to programmatically marked up with SSML and explicitly told to switch to a French model part way through the sentence.

 <speak version="1.0" xml:lang="en-GB" xmlns:mstts="http://www.w3.org/2001/mstts" xmlns:emo="http://www.w3.org/2009/10/emotionml" xmlns="http://www.w3.org/2001/10/synthesis">
 <voice name="en-GB-AdaMultilingualNeural">What is the meaning of </voice>
 <voice name="fr-FR-BrigitteNeural" xml:lang="fr-FR">un sous-sol</voice>
 <voice name="en-GB-AdaMultilingualNeural">?</voice>
 </speak>
๐Ÿ’ก

AI in Education

๐Ÿ’ก

AI in Education

โ€œ
Judge a man by his questions rather than his answers.
Voltaire

The beauty of the LLMs is that they've read everything, if you can ask the right questions >90% of the time they'll give a great answer. Asking the right questions is easier said than done though.

A good prompt to an LLM often requires far more context than a learner can be bothered to give. Successful AI tutors will operate within limits and workflows set by structured data - i.e. the curriculum.

For example in our voice activated vocab builder at Arti ย the OpenAI API is fed data about the learner and curriculum in addition to the user's transcribed response. The LLM uses that data to more accurately provide personalized feedback.

๐Ÿ’ฅ

OpenAI API

๐Ÿ’ฅ

OpenAI API

โ€œ
The true delight is in the finding out rather than in the knowing.
Isaac Asimov

I utilized the OpenAI API as a reasoning engine to help parse and categorise user responses as part of a voice-controlled vocab builder for language learning.

๐Ÿ“ˆ

LLM Fine Tuning

๐Ÿ“ˆ

LLM Fine Tuning

โ€œ
They've done studies, you know. 60% of the time, it works every time.
Brian Fantana

Like everyone else, I'm scrabbling around with LLMs and seeing what sticks! In my work at Arti we tried a number of approaches to improve the output of an LLM, from PEFT to simple prompt-engineering and one-shot inference.

When working with etymological data (about the roots of words) the most effective method we found was simply sense checking the answers against pre-existing structured data, and re-prompting if an error was made.

Skills

OpenAI API

NLP

LLMs

AI in Games

AI in Education

Has your Roomba gained sentience?

If you want to chat about an A.I. project โ€” email me on wjbourchier@gmail.com

Or let's connect: Will Bourchier | LinkedIn