About me
I'm Stewart, or discostew online. I work with local AI models, coding agents, and speech-to-text. This is where I write about what I try and what works.
My faith in Christ is the foundation underneath the work. The goal is simple: build useful things, tell the truth about what works, and help make technology more reliable and accessible.
Things I'm working on
These are personal projects and experiments, built with AI-assisted development. The links go into the code, measurements, and decisions.
oMLX contributions
I tested the components of a larger optimization separately to find which changes justified their complexity. The evidence changed the maintainer's recommendation; three focused PRs were independently benchmarked and merged.
Five contributions are merged overall. A later ablation of another contributor's proposal is submitted as PR #3315, with the original author credited. It remains open.
Local Model Capability Lab
I use images of math problems to investigate what local models can solve. I chose correctness, wall time, and total generated tokens to compare successful answers with the actual wait required to get them.
I inspect available thinking traces for signs of overthinking. Longer output alone does not establish it, and a small problem set does not establish a general model ranking.
Windows PTT Dictation
I use dictation to do my work. Differences in transcription accuracy led me to build a Windows tool with replaceable speech-to-text engines, so I can try newer models without rebuilding the dictation workflow.
The current app uses local Parakeet for live text and supports optional Qwen final transcription on compatible NVIDIA systems. Focus changes, cancellation, and transcript recovery are part of the work, because a useful transcript still needs to reach the right editor.
Groundball
A baseball question-answering app. Ask it something like "Who had the most RBIs in 1962?" and it answers with real numbers and links showing where each answer came from. AI can figure out what to look up, but it doesn't write the factual answer. The app builds that from checked data. Its local statistics engine also runs fully offline, with no AI.
Before the local AI experiments
I spent more than ten years at Microsoft as an Exchange Server Escalation Engineer, from 2000 to 2011.
I investigated complex failures and performance problems using WinDbg, PerfMon, diagnostic and code-level logging, and packet analysis. I reproduced failures, isolated root causes, and prepared defect reports and fix requests for support engineering and product teams.
That work included Exchange public folder replication across databases and servers, networking investigations, and Active Directory issues. Earlier Exchange migrations provided a bridge into the deeper escalation work.
Elsewhere
Why accessibility is personal
Accessibility is personal for me. Years of health challenges changed how I work, and AI tools now let me operate with focus and precision in ways that were not physically possible before. That gives me a direct stake in building workflows that reduce friction instead of creating more of it.
