Skip to content
View wolfvswhale's full-sized avatar

Block or report wolfvswhale

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
wolfvswhale/README.md

J. Alderman Lyell

I write, and I ship software. The two are the same job more often than people expect.

Most of what's here started as a problem in a business. A radon inspection company needed inspection reports that didn't take an afternoon each, so I built a desktop app. A short-form video series needed a production pipeline that could hold a character's face consistent across a hundred episodes, so I built that. I'm very interested in AI detection, quality gates, and AI in general.

What's here

Does a machine-text detector get worse as the generator gets newer? Everyone assumes so. Across eighteen generators released between 2019 and 2025, it does not: within a fixed corpus the correlation runs the other way. What does predict detectability is instruction tuning, by 26 to 29 points on same-family, same-date model pairs.

A quality gate for prose. Eighteen gates, every threshold set by measuring the statistic across 2,602 human-written documents so the false-positive rate is a measured 5% rather than a guess. Fails 46.9% of machine-written documents and 6.3% of human ones. Nine gates don't discriminate on the available corpus, and the README publishes those numbers too.

The evaluation harness underneath it. Structural cadence carries nearly all the signal; the corpus turned out to have a detokenization artifact separating the classes on 73% of documents; and the fitted model transfers below chance to a new domain. The repo documents that failure rather than reporting the in-domain score alone.

Three live business systems: a lead-scoring CRM with auto-routing, milestone invoicing synced to QuickBooks Online, and a legacy data migration that turned a 113-row export into 89 clean typed records while preserving 21 that naive deduplication would have dropped. Screenshots and a reconciled invoice, not mockups.

The generative video system behind a running short-form series. Remotion, Replicate, ElevenLabs, Python, Airtable for production tracking. Sanitized: it's a live commercial series under a pen name, so the persona assets and working tree stay private.

Elsewhere

I write up what I find at wolfvswhale.github.io — short, evidence-first reviews of how machine-text detection gets evaluated, each one linking the code and data behind it.

Long-form historical documentary on YouTube, where I do the research, the script, the edit, and the publish. Books on Amazon under two names. Four years as a digital court reporter before any of this, which is where I learned that a transcript is a product and accuracy is the whole product.

Working with

Python, JavaScript, Node. Claude Code and agentic tooling daily. n8n and Airtable for automation. Remotion, Replicate, ElevenLabs for generative media. React Native and EAS for the iOS side.

Open to remote work. Best reached at jesselyell1@gmail.com

Pinned Loading

  1. vintage-study vintage-study Public

    Does detector accuracy fall as the generator gets newer? Eighteen generators, 2019 to 2025. It does not. Instruction tuning predicts detectability; release date does not.

    Python

  2. bluepencil bluepencil Public

    A quality gate for prose. Fails 46.9% of machine-written documents and 6.3% of human-written ones, with every threshold calibrated against human writing rather than guessed.

    Python

  3. prose-eval prose-eval Public

    Evaluation harness for an editorial rubric that flags machine-sounding prose. Measures corpus leakage, compares rule vs structural features, reports out-of-domain collapse.

    Python

  4. airtable-automation-portfolio airtable-automation-portfolio Public

    Four live business systems, documented as case studies: an Airtable CRM with scripted lead routing, QuickBooks-reconciled milestone invoicing, a zero-dependency data migration pipeline, and an AI v…

    JavaScript

  5. ai-persona-video-pipeline ai-persona-video-pipeline Public

    End-to-end generative video pipeline behind a running AI-persona short-form series — Remotion (React), Replicate, ElevenLabs, Python, Airtable production tracking. Sanitized engineering showcase.

    Python