Company

System: Assembler's Factory Floor Time-Tracking App

Client: Whippendell Marine

Technologies: Raku, CRO, Template::Nest, jQuery, CSS, MySQL

Problem: Create an application that interfaces with the company's existing Teamwork-based project management system and enables assemblers to log time spent on jobs - with design according to client spec.

Solution: A set of "Element" modules were created in Raku which inherited from a base module responsible for making calls to the Teamwork API. Syncing of local application data was accomplished with an independent daemon looping over and caching necessary user and project information in a MySQL database. Server-side rendering created dynamic content with Template::Nest, and the front end assembled HTML chunks delivered from the backend with jQuery. The system was designed to be deployed to Android-based tablet as a PWA app.

System:  Specific details are protected by NDA. However, this system - developed over a number of years - involved creation of a unique architecture utilising LLM technology in a novel way. 

Client: NDA Protected

Technologies: Llama3-70B (and other LLM and ML models), vLLM, Python, Transformers, Algorithms

Problem: NDA Protected - but relates to asset price prediction

Solution: The innovative approach involved making an analogy between the "fuzzy" system we needed to model, and certain more rigorously defined mathematical/physical systems. Specific details are NDA protected. For the implementation we took advantage of per-hour cloud GPU rental to perform the necessary heavy processing as cost-efficiently as possible. 

System: Distributed web-automation application simplifying tasks on a public sector ecommerce portal

Client: MePATool

Technologies: Perl, Javascript, Mojolicious, Hypnotoad, SQL

Problem: The workflow required for public sector ecommerce in a particular jurisdiction can slow down business.

Solution: 7 subsystems have been created and are maintained by Virtual Blue, including client executables, API, Web Admin and web-automation processes. Multithreaded headless Chrome instances are controlled via a custom process management system. Rendering browsers tend to be memory hungry which is one reason the system is distributed across several servers, in order to cope with traffic spikes. Recently very low error rates have been achieved.