Disabled by Design

I design from the standpoint of a disabled trans ethnic studies professor on contingent faculty lines. Every one of those positions is a parameter — shaping what I perceive, what I build, and what I refuse to accept as given. They let me see things somebody with a different neurotype, a different body, a different institutional position would not have thought of. That's a feature, not a deficit.

The name also names the problem: systems are designed in ways that produce disability. Rigid deadlines, surveillance grading, deficit-model assessments — the infrastructure disables people, then calls the people broken. The social model of disability applied to every structure I study and build.

And there's the third meaning: against the impulse to genetic-engineer away disability. I don't want a future that optimizes me out. My existence is an asset.

I don't just build software that accomplishes a task. I reinvent the concept behind the task — interrogate the power relations embedded in it, critique them, and build something that operates from different assumptions.

In my teaching, I reinvented what an assignment could be — from an endpoint for evaluation to a launchpad for discussion. That conceptual shift made the work automatable in the small ways I needed it to be, and the automation opened up the pedagogy. Autograder4Canvas reads student work with a local LLM and surfaces patterns for me, the teacher; it doesn't grade. The prompts ask the model to describe what each student seems to be reaching for rather than score against a static rubric, to attempt to flag silencing patterns like respectability policing or recentering of whiteness, and to attempt to surface burnout signals. They also instruct the model to counter common LLM biases against minoritized students. Whether the model holds any of those instructions on a given run is an empirical question, not a guarantee. FERPA-safe; runs locally on a teacher's laptop. Each piece feeds the other.

I built one tool to help me do more praxis, and it helped me build prompt scaffolding for more praxis. Reframe wraps the LLM call in a 5-stage workflow that asks the model to engage with critical-theory framework specifications before responding — 76 framework specifications, each authored to instruct the model to speak from a specific tradition before synthesis, surface tensions without flattening, and ask whose erasure is a structural condition of the conversation (extending Gayatri Spivak — probably impossible at the prompt-time layer, but worth the attempt). Of the 76, 8 currently have authored detection signatures the system can use to track engagement; the rest are specifications without instrumentation. The session prompts are designed for how I actually think: STEP for parallel infodump-style threads, PARK for tangents, WEAVE for through-lines across a session. A distributed intelligence prompt scaffold attempts to surface through-lines across sessions; that layer is partially implemented.

The critique and the code are the same practice.

Every tool I've built emerged from conditions that demanded it. Autograder4Canvas: built when I was teaching ~200 students across four sections without a TA, to read student work and surface patterns for me. EvalEye: built when the college wasn't processing student evaluations and the forms sat in boxes. RoboStripper: scholarly reading made accessible as audio when the ability to focus on reading degrades under political crisis. Reframe: prompt scaffolding for critical analysis under conditions of teaching under fascism. The interactions with the LLM were going to happen anyway; the question was whether the prompts could route them through critical theory before the model answered.

The conditions that make me precarious are the same conditions that made this system necessary and possible.