johncollins.ai | large-scale deployment of AI: deep learning, LLMs, GPTs

I am an AI specialist focused on large-scale AI deployment, particularly in deep learning and large language models (LLMs). My expertise lies in creating enterprise-level AI solutions that leverage cloud services, data processing, natural language processing (NLP), and robust software design. I address business challenges by establishing PaaS/SaaS AI services with high availability and aligning AI solutions to business needs and trends.

For example, I took Identif.AI from idea to launched product with clients onboarded in less than a year. Identif.AI is an award winning AI-first system engineered around deep learning, LLMs, real-time data streaming, and exceptional UI/UX.

On this site I share insights into my AI work:

project-manager.ai shows the practicality of tailored AI project management frameworks. It combines multiple models, both open-source and closed-source, for flexibility and robustness.
I built a custom GPT, AI Project Navigator, with OpenAI's GPT Builder. The recent unveiling of GPT Builder marks another significant milestone in the usability of GPTs, introducing an emerging layer of abstraction.
I emphasize data privacy in AI, detailing the use of 'internal' models like Llama-2-7B when IP protection is required.
I fine-tuned GPT models for stock price and volatility predictions, comparing prompt engineering to fine-tuning. This highlighted ChatGPT's Advanced Data Analysis advantages.
I addressed enterprise AI compatibility with information security standards. I provided a detailed security questionnaire for AI due diligence.
I implemented Llama-2 as an internal AI assistant, balancing open and closed-source approaches. I showed hands-on implementation on a local machine.
I explored ChatGPT plugins for complex tasks like options pricing. This demonstrated how LLMs can streamline processes into efficient operations.
I developed an AI assistant for due diligence, combining methodologies. This illustrated the rapid advancement of AI models and applications.

My methodical approach to problem-solving, commitment to aligning AI with business, and innovative use of AI for predictive analytics position me at the forefront of the field.

My approach to software development is influenced by AI collaboration. Instead of just coding, I engage in pair programming with AI. This enhances creativity with AI’s expansive knowledge and speed. Key AI enhancements in my workflow include:

Automated code reviews
Code optimization and refactoring
Dependency analysis
Automated testing and prioritization
UI/UX recommendations
Security analysis
Dynamic documentation

You can find source code for anything you see here on my GitHub, and if you would like any further information, please feel free to DM me via the links at top right of this page.

My GitHub

LLMs & GPTs

Working with multiple AI assistants.

AI platforms are quickly evolving into fundamental infrastructure. Moving forward, I believe AI assistants will be central to everyone's engagement with the digital realm.

Through my ongoing project project-manager.ai, I aim to showcase how tailored AI project management frameworks can be obtained from a range of AI assistants. The project serves as a real-life example demonstrating that working with a range of AI models may currently be the most sensible strategy for AI initiatives.

Model Integration: At the core of project-manager.ai, multiple models are combined, illustrating how it is possible to blend open-source and closed-source strategies. This approach mitigates the risks associated with single model dependencies and allows for a richer interactive experience.

project-manager.ai

Flexibility and Context-Awareness: The system uses engineered prompts that adapt to different models and project types, providing context-specific guidance, which underlines the advantages of a combined model strategy.

Robustness: Security and performance optimizations like API leverage, caching, and rate-limiting serve as the backend mechanisms, underscoring the reliability of a combined open-source and closed-source approach.

You may explore the functionalities of project-manager.ai by following the steps below. First, choose an AI-Analyst model from the available options. Next, select the type of project you are interested in. Finally, input a brief description of your project as a prompt. To access the online version, please click on the appropriate links or icons located in the navigation bar.

Custom GPTs

Building a Custom GPT with OpenAI's new GPT Builder

The recent unveiling of GPT Builder marks another significant milestone in the usability of GPTs, introducing an emerging layer of abstraction. I think this inevitably means an expansion in the number and diversity of developers engaging with GPTs, and paves the way for a broad spectrum of applications. I expect that soon advanced custom GPTs will possess an array of capabilities, including reading, writing, auditory processing, verbal communication, visual recognition, artistic creation, and cognitive processing.

GPT Builder enables these models to leverage existing digital tools, specialize in specific domains, access and utilize custom datasets, and perform various actions within the digital landscape. The ability for these models to exhibit tailored communication styles, perform distinct actions, and collaborate with one another, opens up new frontiers in AI.

This advancement is not just a step forward in AI technology; it represents a profound shift in how we interact with and utilize computational resources. As we accelerate into this journey, it's important for stakeholders across industries to prepare for the transformative changes that these technologies are poised to bring.

Here is my Custom GPT, which is called AI Project Navigator. Building and publishing it took less than 30 minutes. Please give it a try by clicking the link, which requires an active ChatGPT Plus subscription for access.

Llama-2-7B

Llama-2-7B can be used as an 'internal' AI assistant when, for example, client data and IP must be protected, alongside 'external' AI assistants such as GPT-4.

Many enterprises I engage with express hesitancy about transmitting their data through public APIs when leveraging AI models like ChatGPT or PaLM-2. The primary apprehensions are rooted in safeguarding intellectual property (IP) and ensuring stringent data privacy protocols. While the latest version of ChatGPT Enterprise does offer robust security measures, including end-to-end encryption and explicit assurances that the user data won't contribute to the ongoing training of OpenAI's models, some institutions may still have specialized requirements that necessitate an isolated environment for model deployment.

In such scenarios, implementing and perhaps fine-tuning an on-premises AI model like Llama-2 can offer a tailored solution.

By adopting this approach, data flow remains confined within the organization's secure network perimeter, and the unique advantages of customizing the model for domain-specific tasks are retained exclusively by the enterprise.

Click continue reading below to see how I have implemented llama-2-7b alongside gpt-4-0613 so that each can be used to produce a basic due diligence report for a given corporate name according to specific circumstances.

Fine-tuning

Leveraging LLMs for time series forecasting of stock prices and volatility.

I take a deep dive into the fine-tuning of language models. I show that fine-tuning is a strong approach for times series prediction. The ability to train a model on specific data, such as price sequences, enhances its understanding and predictive power for those unique use cases.

I fine-tuned a GPT-3.5 Turbo model on 100,000 lines of Apple stock OHLC data to predict future price movements. OpenAI's dashboard for my fine-tuning job summary for the GPT-3.5 Turbo model can be seen to the left.

Fine-Tuning a GPT Model

Rigorously cleasing and preparing high frequency OHLC data for AAPL, the details of which can be found in one of my previous posts: High Frequency Data
Transforming the prepared data into a conversation-style JSONL format suitable for fine-tuning OpenAI's GPT-3.5 Turbo model.
Setting up and initiating a fine-tuning process with OpenAI's API, first by uploading the transformed dataset and then by initiating a fine-tuning job using the uploaded dataset's ID.
Monitoring training until the job succeeded
On the left we see the model ID is: ft:gpt-3.5-turbo-0613:ftiai::8C0RdyUs
Using this model ID with the OpenAI API to generate predictions
Passing the model ID to openai.Completion.create()
Giving it a stock price sequence prompt
Obtaining a prediction for n^steps ahead

Fine-tuning specializes the model for time series forecasting using the stock price dataset

The model is trained to process historical prices and predict the succeeding price accurately. The resulting fine-tuned model is significantly better at this specific task compared to the original, general-purpose model.

Prompt Engineering with Advanced Data Analysis

For comparison, I also experimented with prompt engineering and found that even with quite intricate prompt design, the model's "predictions", in the form of a "hypothetical continuation" appear to be no better than random, at best. Switching to the Advanced Data Analysis plugin I obtained better results. ChatGPT's Advanced Data Analysis does the following:

Data Loading & Extraction: Loaded the provided CSV file and extracted the AAPL_rr and AAPL_lr columns.
Visual Analysis: Plotted both time series to get a visual understanding of their patterns.
Stationarity Test: Used the Augmented Dickey-Fuller (ADF) test to check for stationarity in both series.
Time Series Forecasting: Chose and fitted the ARIMA model and forecasted the next 50 values for each series.
Data Export: Saved the forecasted values for both series to a CSV file.

I would argue that this is a form of advanced prompt engineering where I use components of the broader ChatGPT ecosystem, in this case, to engineer the prompt for us and, in this case, construct the ARIMA model and confront the data to it, which is done on the fly in a Python environment directly constructed and controlled by the LLM.

Prompt Evolution

Evolving Project Management Prompts for AI Projects.

This project demonstrates Prompt Evolution (Fernando et al., 2023) by generating project management prompts tailored to AI projects.

Prompt Evolution uses an iterative process to evolve better prompts for a domain, in this case project management of AI initiatives. As described in "Promptbreeder: Self-Referential Self-Improvement via Prompt Evolution" (Fernando et al., 2023), it leverages a language model to generate variations of prompts, evaluates their utility, and selects the most useful prompts to propagate to the next generation.

Prompt Engineering

Prompt engineering involves a deep understanding of the model's capabilities, idiosyncrasies, and limitations, making it a blend of both science and art.

Prompt engineering involves crafting and optimizing input prompts to guide language models, to produce desired outputs. Beyond simply iterating on prompt structures, methods include prefix-tuning, few-shot and many-shot prompting, templating, and employing meta-prompts, as well as leveraging external knowledge or combining prompts with model fine-tuning to achieve nuanced responses

Prompt engineering to obtain business-specific responses

I built the AI chatbot or 'AI due diligence assistant' below to show how prompt engineering is used to obtain a business-specific response, in this case a due diligence report on a specified company. To use it, enter a company name, select 'Due Diligence' as your AI-Analyst (this is the default for this demo and the button should be green), wait for up to 2 minutes while the model does its work, copy and paste the response into a report and/or send it to yourself via email. The other buttons or response modalities are active, but I haven't worked them extensively yet.