Jonathan – Python, AI, LLM

The project is focused on transforming a private library of WWII classified documentation into a searchable graph and vector database that identifies people, places, and events in 1940s Scandinavia.

• Created a mix of existing libraries and custom algorithms to develop an RAG pipeline for extracting, reconstructing, and storing the information from the archived documents into a set of relational, graph, and vector databases.
• Built a role-based UI with an LLM search assistant (query builder) that searches the databases to return the relevant documents and pages and their related documents.
• Developed a HITL workflow for optimizing the AIs during interactions.

It's a social network that links Hilo community, focusing on video content, live streaming, and local event production.

• Built and deployed the foundational tech stack within a two-week timeframe.
• Leveraged local events to boost traffic and drive user adoption.
• Currently finalizing the CI/CD pipeline using GitLab, Nomad, Docker, Consul, and Traefik.

It's a modern ARM platform that integrates with your tools, automates workflows with AI, and delivers real-time insights through an intuitive, always-current interface.

• Took over the project when the previous AI Engineering team left the company.
• Redesigned existing models with optimized objective functions aligned to maximize ROI.
• Spearheaded the transition from AWS to OCI, ensuring a smooth and efficient migration process.
• Directed a team of 9 engineers through a reduction to 5, ensuring continued project success and team cohesion.

It's a fully offline LLM that I built, which learns from users, incorporates customizable personalities via RAG, and supports household tasks, including IoT device management, calendar, and alerts.

• Engineered a solution to host a 7B-node LLM on outdated hardware.
• Developed and refined four iterations of a custom-built knowledge graph, culminating in a modern RAG system with advanced NLP capabilities.
• Integrated a streaming text-to-vector (T2V) large language model.
• Built an initial version of a custom workflow engine to streamline model and task management.
• Partially completed API orchestration layer to enable seamless integration of additional features.

It's an Insurance Hold Manager that used a combination of IVR Navigation, Audio Fingerprinting, and Regression Neural Network models to increase the productivity of our customers (US Healthcare Billing Followup) by 40%.

• Developed original AI models entirely from scratch without relying on frameworks like PyTorch or TensorFlow.
• Designed and implemented the original audio fingerprinting solution entirely from scratch.
• Co-led the startup through its successful acquisition by another company.
• Led teams ranging from 1 to 14 members across various projects.