About Me
I'm Jillian — a Deaf, neurodivergent, disabled, non‑traditional coder based in Colorado. I build local‑first, community‑first tools using Raspberry Pi systems, AMD ROCm, and open‑source software.
My work centers on accessibility, autonomy, and evidence‑driven engineering. I specialize in assistive‑tech search tools, RDNA3‑safe model training, and resilient offline workflows.
GitHub:
Spoonie Helper
Spoonie Helper 🥄 is an assistive‑technology price comparison tool prioritizing Deaf‑owned, disabled‑owned, and community‑based vendors before major retailers.
Originally part of AccessiFind, it has evolved into a multi‑agent local‑first system using:
- Qwen 2.5 3B for reasoning and ranking
- Phi 3.5 Mini for classification and metadata cleanup
- TinyLlama for background tasks and offline inference
These agents power natural‑language search, accessibility‑aware ranking, and fully offline operation on Raspberry Pi hardware.
Current status (June 2026):
- Backend ingestion + ranking engine complete
- Frontend search interface in progress
- Local Pi agent running continuously
- Accessibility‑aware vendor scoring finalized
Local Compute and Model Training
My home training environment uses a Raspberry Pi 3B+ paired with an AMD ROCm 7.2.1 + PyTorch RX 7700 XT GPU system.
The Pi runs TinyLlama as a lightweight model server and task agent. The GPU system handles RDNA3‑safe QLoRA fine‑tuning of Qwen 2.5 3B, TinyLlama variants, and experimental small models.
This setup supports fully offline workflows, enabling accessible and resilient development. My ROCm 7.2.1 QLoRA pipeline was submitted to the AMD Hardware Challenge.
Radeon Developer Kit
I work with the AMD Radeon Developer Kit as part of the Radeon Developer Program. This hardware provides a stable, well‑documented platform for testing ROCm behavior, validating my QLoRA pipeline, and capturing reproducible GPU‑state evidence.
The dev kit supports:
- RDNA3 inference + kernel‑behavior testing
- ROCm SMI logging + GPU‑state capture
- Cross‑device reproducibility checks
- Stable vendor‑supported development for accessibility‑focused tools
It strengthens my local‑first workflow by giving me a second AMD platform for debugging, validation, and long‑term reproducibility.
AI + Accessibility Research
My current research explores how accessibility principles stabilize AI behavior. I study how AI brittleness mirrors healthcare workflow failures: correct‑looking output, wrong meaning, no flag.
I use Microsoft Copilot, GitHub Copilot, and Copilot Health as assistive technology to reduce physical strain, structure complex reasoning, and translate hardware‑level insights into accessible language.
This work has gained traction among engineers, interpreters, accessibility advocates, and AMD community members — and is informing early policy discussions in Denver.
Read the full RDNA3 + AI Research (new‑research.md)
Read the Healthcare AI + Accessibility Analysis (healthcare-ai.md)
Evidence‑Driven Workflow
I maintain a public evidence archive documenting GPU behavior, system metrics, incremental‑training lineage, validation runs, and accessibility‑related screenshots. This archive mirrors my local development environment and is updated continuously.
The archive includes:
- ROCm SMI snapshots + system‑state logs
- Training summaries for Qwen 2.5 + TinyLlama
- Validation traces + dataset behavior logs
- Accessibility documentation + GPU‑state screenshots
View Appendix‑A: Full Evidence‑Driven Workflow Archive (GitHub)
Contact
Email: jg@jg18.dev