Various Clients | Design Leadership | 2023
By 2026, 80% of enterprises will have used Artificial Intelligence (AI) APIs & models. David and his teams are working to lead that AI experience.
David sees tremendous design opportunities for how companies humanize AI for users in ways that feel intuitive, non-invasive, transparent, and non-creepy. Today, there is a gap between the reality and perception of AI, but this is more about the fundamental context than the potential or value of the technology itself. Even if it’s not all hype when it comes to AI, AI leaders are concerned that they are overpromising and underdelivering. David leads teams to mine the art of the possible to unlock real-world value for users and businesses.
The Cutting-Edge Work
AI has evolved significantly over the past few years and has been applied to almost every business sector. These are some examples of how David led teams in this pursuit.
New York City
A team for NYC Mayor Eric Adams’ Office for Contract Services to develop an AI called MOCSY. MOCSY is leveraging Conversation Engines, AI Support, Advanced UX Assist, Multi-Modal Search, and Testing Via Natural Language. A key KPI will be an 83% reduction in incoming service desk tickets ensuring a decrease in call volumes for the Mayor’s teams.
FreshDirect
A POC demonstrating how predictive modeling can be used to recommend recipes and products related to individual shopping behavior. AI: LangChain to compose queries to GPT-3.5 that generate recipe recommendations based on behavioral analytics and product catalog data.
Design Search
A way to socialize the firm’s collective design history and inspire best practices for a globally distributed design group. Al: OpenAl’s CLIP model utilizing multi-modal search (text and pixels) allows users to quickly find relevant Figma design assets.
A Law Institute
A conversational Ul interface for: 1. Text: Pre-gen training to provide classification and summarization of key themes; 2. Search: Using an LLM to understand the semantics of legal data and clustering verification; 3. Video: Convolutional neural networking to classify facial emotions; and 4. Presentations: creating decks of key themes, descriptions, and deep-links to video content.
TestPilot
A QA testing tool to write and execute functional tests with natural language prompts. AI: GPT-4o API to convert functional tests written in natural language into instructions for an accessibility screen reader to execute.
A Deeper Dive
Fortune | Commerce | Travel
FORTUNE x ACCENTURE: Fortune Analytics
After 95 years, Fortune entered the age of AI with a new tool aimed at better analyzing the Fortune 500. David was a key leader in a team that developed a new product, called Fortune Analytics. It will deliver ChatGPT-style responses based on 20 years of financial data from the Fortune 500 and Global 500 lists. “What it is going to do is take over 20 years of our incredible trusted journalism, combined together with our financial research and analysis and financial data, and put it all together into one easy-to-use tool.” – Anastasia Nyrkovskaya, CEO. Drawing on decades of in-house data, Fortune Analytics will allow customers to build graphic visualizations with the help of an easy-to-use AI assistant.
The platform is powered by Accenture’s foundation model services, a proprietary large language model, and is fine-tuned with comprehensive Fortune datasets. David led the design team in research, POCs, UX/UI and how the platform understands plain language requests and delivers relevant answers.
Commerce UX/UI
David’s team has designed several AI-driven chat experiences for online shopping. This user experience framework combines the power of natural language processing paired with detailed product information to orchestrate an intuitive user interface. How it uses Al: LangChain combines services to determine the intent of user requests and compose queries to the GPT-4o API that return recommendations informed by a database of products and user reviews. The apps personalize every shopper interaction to better engage users and drive sales. This AI: 1. captures every conversation in one place: brings fragmented customer conversations, channels, and tools into one unified helpdesk to increase support efficiency. 2. automates 60%+ of customer support: AI and automation tools are built on a company’s own data and integrations, quickly providing the right answer to their customers every time. And 3. unlocks Customer Lifetime Value: proactively engages with customers at key moments on a company’s website with on-brand campaigns to increase GMV.
AI Travel Planning
David’s team tackled crafting unforgettable itineraries with an AI trip planner. The personal trip planner and travel curator creates custom itineraries tailored to a user’s interests and budget. Also, by automating tasks, providing personalized recommendations, and streamlining decision-making, Al-driven simultaneous agents save time, reduce errors, and create more efficient and tailored plans. Lastly, a user can simply drop a link to their favorite travel content and AI will create a customizable list or itinerary for in seconds.
Want to Know More About AI?
Decoding the relationship between AI, ML, DL, LLMs, and GenAI.
Artificial Intelligence
AI is a discipline. It refers to programs that sense, reason, adapt, and act.
SUMMARY
- An overall, all-encompassing framework for the theory and development of computer systems performing tasks normally requiring human intelligence and cognition.
Machine Learning
ML is a subset of AI. It uses algorithms whose performance improves as they are exposed to more data over time. ML gives computers the ability to learn without being explicitly programmed. Typically, machine learning involves: 1. building or using an existing algorithm; 2. feeding data into an algorithm; 3. using this data to train a model; 4. testing and deployment of the model; and 5. using the model in an app.
Most of the machine learning algorithms are categorized into these learning models: supervised, unsupervised, reinforced, and deep learning.
SUMMARY
- Algorithms whose performance improves as they are exposed to more data over time
- Specific algorithms, historical data processing, time series data
- Gives computers the ability to learn without being explicitly programmed.
Deep Learning
DL is a subset of ML where Multilayered Deep Neural Networks learn from vast amounts of data. These algorithms can learn how to make an accurate prediction through their own data processing, thanks to the artificial neural network structure. Deep Learning models are designed to automatically and adaptively learn from experience using neural networks with multiple layers.
There are two common DL model types: discriminative models and generative models.
SUMMARY
- Subset of Machine Learning in which Multilayered Deep Neural Networks learn from vast amounts of data
- Learning from a broad range of data
Generative AI (GenAi)
GenAI is a subset of DL that uses techniques like DL algorithms or transformers to generate new content.
Typically, the models leverage massive pre-trained knowledge to generate content. Large Language Models are a type of machine learning model for Natural Language Processing (NLP). These models are trained on vast amounts of text data and can generate human-like text (and many more things like images, code, etc.) by predicting the likelihood of a given word sequence.
SUMMARY
- Subset of Deep Learning that uses techniques (such as deep
learning) to generate new content. - Typically, the models leverage massive pre-trained knowledge to generate content.
Diving Into GenAI
Widely accessible GenAI tools can address highly specific use cases and enable a broad spectrum of cross-product experiences. David and his teams have experimented with most available GenAI model / provider layers and have successfully launched various integrations.
TEXT
Capable of ingesting and generating input prompts; input and output text lengths vary by model.
CODE
Impact the software development lifecycle and coding productivity (e.g., GitHub CoPilot).
IMAGES
Capable of generating many aesthetic styles and techniques for editing and modifying; models are subject hallucinations and volatile image quality.
AUDIO
Used to convert between speech and text, generate new sounds, and create music based on an input prompt.
VIDEO & 3D
Potential to unlock creative markets like cinema, gaming, VR, architecture, and physical product design.
OTHER
Fundamental model R&D across many fields from art to drug discovery & theoretical synthesis.
OpenAI GPT-4o,
Anthropic Claude 2,
Google PaLM 2,
Meta LLaMA 2,
Cohere Command,
Falcon
OpenAI GPT-4o,
Code Whisperer AWS,
Code LLaMA,
Codey,
Stable Diffusion,
OpenAI Dall-E 2,
Midjourney,
Imagen,
Runway,
Mosaic PT
OpenAI Whisper,
Meta Seamless M4T,
Stable Audio,
Microsoft X-CLIP,
Meta Make-A-Video,
Gen-2,
Soundify,
DreamFusion,
NVIDIA GET3D,
MDM,
Magic 3D
Galactica,
Sambanova,
Minerva,
Alphatensor,
GATO,
Human Motion Diffusion Model
The Possibilities
Opportunities David sees the most in the market.
CONVERSATION ENGINES
This user experience approach combines the power of natural language processing with detailed product information to intelligently orchestrate an intuitive user interface.
PRE-GENERATIVE TRAINING
Pre-generative training on OpenAl to provide classification and summarization of key themes.
ADVANCED UX ASSIST
A proof of concept demonstrating how predictive modeling recommends directions and products related to an individual’s behavior.
MULTI-MODAL SEARCH
MMSearch is a tool that provides multi-modal search capabilities (between text and images), allowing users to quickly find assets across vast pools of data.
NATURAL LANGUAGE
Convert functional tests written in natural language into instructions for an accessibility screen reader to execute.
LangChain combines services to determine the intent of user requests and compose queries to the GPT-4o API that return recommendations informed by a database of products and user reviews.
Using an LLM to understand the semantics of data for better search and verification.
We leveraged the LangChain composability framework to compose queries to GPT-4o that generate recipe recommendations based behavioral analytics and product catalog data.
We have utilized OpenAl’s CLIP model.
We use GPT-4o API.