Paul Gaynor

Education

Certificate (Machine Learning with Python(6.86x)) M.I.T. (online via edX.) 2020 Show Details

Course Description (from the institution)

Machine learning methods are commonly used across engineering and sciences, from computer systems to physics. Moreover, commercial sites such as search engines, recommender systems, advertisers, and financial institutions employ machine learning algorithms for content recommendation, predicting customer behavior, compliance, or risk. As a discipline, machine learning tries to design and understand computer programs that learn from experience for the purpose of prediction or control.
In this course, you will learn about principles and algorithms for turning training data into effective automated predictions. We will cover:
- Representation, over-fitting, regularization, generalization, VC dimension;
- Clustering, classification, recommender problems, probabilistic modeling, reinforcement learning;
- On-line algorithms, support vector machines, and neural networks/deep learning.
You will be able to:
- Understand principles behind machine learning problems such as classification, regression, clustering, and reinforcement learning
- Implement and analyze models such as linear models, kernel machines, neural networks, and graphical models
- Choose suitable models for different applications
- Implement and organize machine learning projects, from training, validation, parameter tuning, to feature engineering
You will implement and experiment with the algorithms in several Python projects designed for different practical applications.
You will expand your statistical knowledge to not only include a list of methods, but also the mathematical principles that link these methods together, equipping you with the tools you need to develop new ones.

Certificate (Neural Networks for Machine Learning) Univ. of Toronto(online via Coursera) 2018 Show Details

Taught by the legendary Geoffery Hinton, the course covered

The Perceptron learning procedure.
The backpropagation learning procedure.
Learning feature vectors for words.
Object recognition with neural nets.
Optimization: How to make the learning go faster.
Recurrent neural networks.
More recurrent neural networks.
Ways to make neural networks generalize better.
Combining multiple neural networks to improve generalization.
Hopfield nets and Boltzmann machines.
Restricted Boltzmann machines (RBMs).
Stacking RBMs to make Deep Belief Nets.
Deep neural nets with generative pre-training.
Modeling hierarchical structure with neural nets.
Recent applications of deep neural nets.

PhD(Computer Science),The University of the West Indies(UWI), Jamaica. 2016

MSc(Telecommunications Technology), UWI, Jamaica. 2001

BSc-1st Class Hons(Electronics + Computer Science), UWI, Jamaica. 1995

Projects

Digital transformation champion for the Kingston YMCA, responsible for building and deploying cloud-enabled data management app(2023-2024) Show Details

Problem:The (mainly manual) data management operations have been resulting in significant financial loss, and present difficulties in compliance to the recently passed Data Protection Act.
Solution: I built and deployed an internal data management system that supports secure democratization of personal data management using a mobile application.
Results: The office has an online system available to securely manage data, and the app is to be placed in the Google PlayStore shortly.
Impact: Still being assessed. Expected impact includes improved customer retention, as well as additional diversity in the management of activities, and physical space resources.
A point of interest: The opportunity to integrate a recommendation model within the app that drives the most effective ads to users, using local user information.

Developed and deployed the Stack Seven Process Manager that tracks food processor quality to meet US FSCPA import regulations (2023-2024) Show Details

Problem:Food products created for the US import market are required to strictly document processes, as well as batch creation history.
Solution: I built and deployed a process that defines the marketing processes, and allows users to enter activity on each product batch .
Results:Cloud based process flow manager created, along with mobile app to enter batch data.
Impact: Still being assessed..
A point of interest: The opportunity to integrate a micro-LLM that suggests steps, hazards and measures for potential products.

Deployed the Kingston YMCA's website kymca.net (2023) Show Details

Problem:The Kingston YMCA needed a more consistent way to improve it's engagement to stakeholders than social media, and to allow donations.
Solution: I designed and deployed the site that would spread the message of the YMCA, and allow easy content update.
Results: The site has been deployed, and a significant number of donations have been received.
Impact: Still being assessed, as the site is expected to work in conjunction with the app and the data management processed.
A point of interest:The ability to visually tell a coherent story about the operations of the Kingston YMCA through the slide shows.

Implemented a video monitoring prototype with drone response using Flask, REST, Embedded Computer vision and custom AI models.(2021-2023) Show Details

Problem:Manual monitoring of surveillance streams time intensive, error prone, and does not scale.
Solution: Develop an automated monitoring platform that captions and indexes interesting video sequences, using custom AI models based clustering to help assess relevance. Show Details

The aim of the MovingEye project was to build a prototype of a video monitoring solution that could meet the following objectives:

Index interesting activities to allow easy search for later retrieval.
Send alerts to authorized personnel if potentially risky events are likely to be in progress.

The need for such a system arose from the high levels of human effort and attention required to extract actionable intelligence from camera feeds. A such, even with full coverage of an area of interest using surveillance cameras, the impact in minimizing violent crime events is often not realized. Meeting objective 1 required assembling software and hardware components that were readily available at the time of project initiation (mid-2022), writing some code, and fine-tuning some existing AI models that performed video captioning. Objective 2 was more challenging, for several reasons. These include the fact that delays are introduced when transmitting data from video collection points to the places where decisions are made, as well as when processing video, and that the possible set of potentially risky events may vary widely with the context of the video stream. Also, the processing to detect a risky event is often not real time, so the time to fully process an event takes longer than the event. The solution applied to objective 2 was to deploy embedded computing devices as close as possible to the video recording points that selectively look at the events are to be processed, and perform three other tasks, namely captioning, prediction and alerting. The selection method applied was to detect when the context of a video has changed. (An example of when a change is detected is if the stream shows persons sitting and eating in one instant, and then shortly after sees persons running). Given that there was no known dataset that collated how video context changed, the research team created a customized process that grouped and clustered videos (using the Microsoft Research – Video to text dataset). The second task applied was to perform captioning on the video sequence. As such, each caption was considered a part of a story. The prediction aspect was expected to extrapolate the most likely next event in the story. This was important given the delays in processing, as, if the processing was behind time, the only way the process could try to send danger alerts in time would be to try to predict (Note that a prediction is a qualified guess, so there is no guarantee the prediction will be realized). The researcher in charge of this feature fine-tuned a GPT2 model to perform the prediction, which was a cutting edge model in 2022. Finally, if the prediction was classified as a danger (we used a combination of lookups and sentiment analysis from a language processing toolkit), an alert is sent to authorized personnel.

Results: Prototype has been developed.
Impact: Platform to conduct further research is available.
A point of interest:Development of the custom model to assess potential relevance detects contexts change by assessing whether the current frame sequence indicates deviation out of a cluster, where clusters were generated from the Microsoft Research Video to Text dataset..

Built and deployed outdoor media management system for National Outdoor Advertising-Jamaica.(2008-2010, 2015-2018) Show Details

Problem:Management of advertising locations was cumbersome for National Outdoor Advertising.
Solution: I implemented a system that performed automated contracting, custom reporting and marketing .
Results:System deployed using Visual Basic, SQL Server, and Office automation that maintained data consisency for over 50 users.
Impact:Automation of previously manual system to seamlessly manage multimillion dollar annual transactions.
A point of interest:After management insisted on a budget cap but users insisted they required reports outside the budget, I implemented a solution that met both parties using an ad-hoc report feature that allows custom reports to be csreated by authorized users.

Engineered bayesian analysis ETL (Extract, Transform and Load system) using SQL Server, MySql, PHP, HTML5, and the Paypal API.(2014-2015)Show Details

Problem:Proprieter of the Grover Group had an idea for an ad-hoc bayesian analysis platform, and needed someone to build it.
Solution: I implemented a system that can import, sanitize, estimate billing, and apply Bayesian analysis to ad-hoc data sets.
Results:System deployed SQL Server, PHP, HTML5, JQuery, Paypal interface that is able to support over 100 users.
Impact:Implemention of billable product from ideation stage.
A point of interest:This was my most fruitful project from the outsourcing site elance.com. Development initially started over a mysql base but we moved to another database platform due to limited features. The development phase formally ended in 2016, so there is no guarantee that the client is using the same underlying platform.

Computer vision based toll estimator.(2013)Show Details

Problem:Toll system operator wanted to estimate and validate toll cost of vehicles entering booth.
Solution:Apply relevant OpenCV libraries with appropriate heuristics to estimate the size of a vehicle in a frame, then categorize and report.
Results:System delivered to client.
A point of interest:This was my first commercial computer vision based project.

Developed ETL to implement Financial Information eXchange protocol using SQL Server and SSIS.(2012)Show Details

Problem:Jamaica Stock Exchange was updgrading to a new data transport protocol, and the stock broker needed to upate..
Solution:I encoded FIX business rules using robust high transaction strategies that leveraged database consistency.
Results:Reliable solution delivered .
Impact:Capability of live data updates within trading day (upgraded from previous end of day update).
A point of interest:The planned 3 month schedule was challenging but not impossible. A month into the project, the schedule was effectively slashed by a month. I bit the bullet and delivered within the updated timeline anyway.

Designed, built and deployed a Cambio (Foreign Exchange) Management system (2005-2006).Show Details

Problem:Integrating JMMB's cross currency transactions was cumbersome, with challenges in managing daily close rates across branches.
Solution: I refined the design and deployed a server-based FX trading solution based in SQL logic with interfaces to relevant systems .
Results:Reliable solution delivered using PHP, Sql and Crystal Reports.
Impact:Automation of previously manual system to seamlessly manage daily multimillion dollar transactions.
A point of interest:Existing asynchronous operations across branches created a challenge in the daily closeout process which was a required point of synchronization. The design therefore required that although each branch could mark the end of daily trading, all consistency checks for all branches need confirmation before the system rolls to the next day, with escalatable alerts to authorized personnel if the event of bottlenecks.

Created a ship pilot scheduler for the Port Authority of Jamaica using .Net and SQL Server.(2004-2005)Show Details

Problem:The Port Authority of Jamaica required a transparent strategy to schedule licensed ship captains to bring ships into Jamaica's ports.
Solution:Built and deployed Pilot Dispatch System that generated fair dispatch schedules for local ship captains
Results:Reliable solution delivered over SQL Server and C# (verified to be still operational in 2020).
Impact:Availability of transparent reliable strategy to schedule Jamaica's ship captains.

Developed and supported equity stock management system for JMMB Securities Ltd.(2002-2007)Show Details

Problem:An equity stock trading system requires accurate, timely information with efficient execution of trade decisions.
Solution:Create online stock ordering, approval, fulfillment, contracting and notification system.
Results:Fast, reliable system delivered over SQL Server, SSIS, ASP, PHP, JavaScript,Crystal, Platinum.
Impact:Automation of previously manual system to seamlessly manage multibillion dollar transactions that supported over 10,000 users.
A point of interest:The sheer scope of the project, starting from manual operations was an interesting undertaking. Many other notable features were also encountered, from the design that allowed cross layer reporting of error conditions along with extensive coupling across applications, to the dynamic alert system, which was novel at the time.

Designed and developed Math teaching application for the Ministry of Education(MOE), Jamaica (ROSE project)::.(2000)Show Details

Problem:The MOE wanted to improve math teaching capability, and acquired funding to build a system.
Solution:I coordinated a team of teachers and developers to encode lessons and assessments a distributed application with HTML and Flash.
Results:System tested and approved by MOE, was to be deployed in a follow-up "ROSE 2" project.
Impact:Useful artifact created that showcased capability of online learning.
A point of interest:Starting from nothing and with consideration of respect for existing intellectual property, a complete set of project activities that include identifying and recruiting teachers and developers, as well as getting tasks executed and ensuring quality. To think, all this was done before I understood the concept of PMP certification, giving me valuable on the job training!!!

Designed, developed and deployed resettlement manager for the North Coast Highway Improvement Project (NCHIP), Ministry of Transport(1999)Show Details

Problem:A pre-requisite manadated by the IDB for the NCHIP project was documentation of the compensation to all project-affected persons (PAPs).
Solution:I identified disparate field agents' data, then sanitized and normalized it to create unified reporting over SQL Server.
Results:Centralized system deployed from the previous attempt to maintain inconsistent silos.
Impact:100% PAP compensation.
Three points of interest:
- This was my first national scale project, and the largest to date.
- In earning this contract, I had needed to hunt for work given the situation at the time. My mindet was to start Telecommunications postgraduate program "or bust" within five years of completing my undergrad. A program had just been started at the University of the West Indies and I did not have funds to finance it. So, I took out a bank loan to get started on the first semester, with no clue where the rest of the funds would come from. Proceeds from the project were required to clear the loan and complete payment for the tuition.
- The reputation earned while executing this project was the springboard for subsequent consultancy assignments, paving the way for me to run my own software development company for several years.

Experience

Education

Resume

Projects

Current Certifications

Prior Certifications

Community Outreach and Leadership

Contact Me