Keith Hackworth

Contributed to establishing engineering practices and scaled enterprise delivery capabilities for a platform supporting nearly 20,000 active developers and 530+ production applications. Co-founded a DevOps organization and helped define build/release, CI/CD, and software-delivery practices across the platform. Authored an enterprise governance framework with automated validation and led remediation across many application teams, reducing outstanding findings substantially.

Founded an independent R&D and product development company and built an AI-driven interactive environment focused on accessibility and entertainment. Implemented local LLM pipelines and multi-agent workflows, including LoRA fine-tuning and AI control logic that orchestrates C++ outputs for ESP32 lighting and effects. Built an ASL recognition model using temporal MediaPipe landmarks and developed additional products including an organizational intelligence platform using RAG and MCP with a MongoDB knowledge base.

Operated a Credit Murex UAT environment by rebuilding and configuring servers while supporting frequent release cycles. Managed U.S. and Singapore production support teams covering UAT environments, server operations, and global production activities. Built an automation platform for environment provisioning, configuration, application restarts, test-data refresh, batch end-of-day processing, and production deployments.

Served as a primary systems administrator supporting servers, enterprise web properties, and company technology operations, later rising to acting associate director and second-in-command of a large technology organization. Architected and built an enterprise onboarding and provisioning platform and designed enterprise data center deployments with production migrations maintaining high availability. Managed a multi-million-dollar operating budget and renegotiated hosting agreements while improving service levels.

Provided support for development, integration, deployment, and operations of command, control, and tracking capabilities used by military organizations and the DoD. Worked across shipboard and installation environments, troubleshooting system boundary issues to maintain operational availability. Coordinated with military personnel and technical teams to diagnose field issues and support deployed systems, while holding a U.S. Top Secret clearance with SCI processing initiated but not completed.

HandFlow is an open source reusable TypeScript library for turning frame-by-frame hand-tracking data into persistent, time-aware information that applications can reason about across a video stream.

Built to sit above computer-vision frameworks such as MediaPipe, HandFlow maintains hand and person identity across frames while tracking landmark history, timestamps, movement and velocity. It also provides handedness confirmation and configurable frame history/delay, allowing applications to analyze how a hand moves over time rather than treating each camera frame as an independent observation.

I originally developed HandFlow while working on real-time ASL recognition, where individual hand poses were not sufficient to distinguish signs involving movement and transitions. I separated the tracking and temporal-processing functionality into its own library so it could be reused for gesture recognition, motion analysis, accessibility interfaces, and other computer-vision applications.

The library is written in TypeScript with automated tests covering its core tracking and history functionality.

https://github.com/keithhackworth/handflow

N-Dimensional Machine Learning (NDML) is an experimental machine-learning approach I developed to explore classification based on stable relationships between multiple dimensions of observed data rather than relying solely on individual feature values. This is critical to fill a gap in traditional transformer models, where it doesn't understand relationships over temporal datasets, which is critical for my ASL project.

The system models relationships across dimensions and over time, allowing classes to remain independent while identifying patterns that persist despite variation in individual observations. I implemented a two-pass coarse-to-fine classification approach designed to quickly identify likely candidates before performing more detailed comparisons.

Built the experimental framework, training and classification pipeline, data-processing tools, and metrics needed to evaluate accuracy and behavior across different configurations. I used the project to test the approach against real-world hand-tracking data and investigate how temporal and multidimensional relationships could improve recognition of gestures and other dynamic patterns.

The research was later paused after identifying inconsistencies in the underlying MediaPipe landmark data that made it difficult to distinguish limitations in the NDML approach from variability introduced by the input data.

SonicBooth is a custom-built interactive AI environment combining software, AI, computer vision, embedded hardware, sensors, cameras, lighting, audio, fans, and physical controls into a single adaptive platform. I designed and built the system end-to-end, including the application architecture, React-based interfaces, Python services, REST APIs, ESP32 and Orange Pi integrations, custom electronics, and communication between physical and software components.

The platform supports multiple adaptive AI personalities and experiments with natural interaction through voice, vision, eye tracking, gestures, environmental sensing, and physical controls. Accessibility is a core design goal, with alternative interaction methods designed for users with visual, hearing, speech, or hand limitations.

SonicBooth also serves as my personal R&D platform for experimenting with LLMs, AI agents, RAG, MCP, computer vision, custom machine-learning approaches, and new human-computer interaction concepts.

Experimental real-time American Sign Language recognition system using computer vision and machine learning to identify signs from hand position, shape, movement, and changes over time.

Built a vision pipeline using MediaPipe Hand Landmarker, OpenCV, Python, and TensorFlow Lite to capture and normalize hand landmarks from live video. The system analyzes sequences of frames rather than relying only on static poses, allowing signs to incorporate hand movement, direction, velocity, handedness, and temporal relationships.

Developed custom data collection, training, testing, and recognition tooling while experimenting with approaches for distinguishing visually similar signs and handling natural variation between users. The project also led to development of HandFlow, a reusable TypeScript library for maintaining hand-tracking history, timestamps, velocity, handedness confirmation, and person-level grouping.

The work is part of a broader effort to explore accessible, natural human-computer interaction for SonicBooth.

KnowMy.biz is an experimental organizational-intelligence platform built around a simple problem: a business doesn't always know what it knows.

The system uses adaptive AI-driven interviews to learn how an organization actually operates from the people who work within it. Rather than following a fixed questionnaire, it builds a persistent model of the business—including people, roles, responsibilities, processes, systems, dependencies, exceptions, problems, and institutional knowledge—and uses what it has already learned to determine what questions should be asked next.

I designed the platform around a persistent MongoDB knowledge base with evidence and provenance, allowing information from different employees and sources to be reconciled while preserving uncertainty, contradictions, and unanswered questions rather than simply treating every statement as fact.

RAG and MCP provide additional context and investigative capabilities, allowing AI agents to search existing organizational knowledge, identify gaps or conflicting information, perform deeper investigations, and generate targeted follow-up questions.

The long-term goal is to create a living model of how a business actually works—not just how its documentation says it works—so the system can preserve institutional knowledge, expose dependencies and inconsistencies, answer questions about the organization, and eventually identify opportunities for improvement and automation.

I put this project on hold for a while. I had a "eureka!" moment for my longer term ASL project and switched gears.

Demo: [email protected] / demouser

I built the majority of Live-DDL in less than a week, with additional features and refinements added over the following week. I needed a faster way to develop database schemas and ran out of free credits on the online tools I was using, so I built my own.

Live-DDL is a visual database design platform that keeps SQL DDL, interactive entity-relationship diagrams, and database metadata synchronized in real time. Developers can work directly with SQL, visually manipulate the schema, or edit table, column, index, and relationship properties without the different representations drifting out of sync.

I designed and built the platform around a database-neutral schema model with bidirectional parsing and generation of SQL DDL. It supports complex relational structures including composite keys and indexes, self-referencing and multi-column foreign keys, referential actions, and conceptual relationships that document application-enforced relationships without requiring physical database constraints.

The interactive ERD provides automatic layout and relationship routing while preserving manually arranged layouts and connector paths. Larger schemas can be explored by focusing on selected portions of the model without losing the context of the complete database.

Live-DDL also includes project persistence, revision history, authentication, permissions, organizational workspaces, and multi-user project support. The architecture is designed to support multiple database platforms and provide a foundation for schema validation, standards enforcement, database synchronization, and AI-assisted database design.