McNeillium_AI: A Checked Video Pipeline
Claude writes, ElevenLabs narrates, Python and FFmpeg assemble, and nothing ships until every fact is traced and every clip is transcribed
- The problem
- Automated video makes it easy to publish claims nobody has checked.
- My contribution
- Built a Python pipeline from script to captioned video, with speech-to-text and source checks before anything ships.
- What you can inspect
- Public repository and case study. The six explainers are built but not yet published.
How it fits together
Overview
McNeillium_AI started in May 2026 as an experiment: how much of making a short explainer video can be automated without losing control of what it says? It grew into a pipeline of about 24,000 lines of Python with 43 defined roles, from trend research and scriptwriting to sound design and quality review, which Claude takes on in turn under rules I set.
Its most useful output so far is a series of six short explainers of my own work, filmed from this site.
How it works
- Research and script. A researcher role scores candidate topics; Claude writes a structured script with segments, narration and shots; a reviewer role checks it against a checklist.
- Narration. Each segment is voiced separately (ElevenLabs, or Edge TTS for drafts), then checked automatically for audio quality and loudness.
- Visuals. Shots are screenshots of real pages taken with Playwright, images from open libraries, icons and generated diagrams, chosen by a “visual director” step.
- Assembly. FFmpeg renders each scene, joins them, times captions to the speech word by word and burns them in, and makes a thumbnail.
- Packaging. An SRT caption file, YouTube metadata and a private upload. A cost governor refuses paid API calls once a daily or monthly spending limit is reached.
The explainer series
Six videos of 1 to 2 minutes each: Voice, the river study, the drug-response study, six rules for checkable research, the portfolio’s narration and voice assistant, and the Lab demos. The rules for this series are stricter than for anything else the pipeline makes:
- Every fact must come from this site or a public repository. Research videos say the work has not been peer reviewed.
- The narrator is an AI voice, speaks in the third person, and every description says it is not my voice.
- Each clip is transcribed with Whisper and compared with its script, and the frames are reviewed, before packaging.
- Uploads are private first; I watch each one before anything is published.
Checking caught two errors before recording: a figure from the drug study quoted for the wrong drug, and a river experiment described as re-run when it had not been yet.
All six are below, hosted on this site.
-
Voice: an AAC app that explains its predictions
Transcript
This is Voice, an AAC app built by Paul McNeill. AAC stands for augmentative and alternative communication: a way to talk for people who cannot rely on speech.
The user builds a message by tapping words on a board of colour-coded tiles. Every word keeps the same place, in Child and Adult modes alike, so the board can be learned by position instead of searched.
As the message grows, Voice suggests the next word. Each suggestion says why it was offered, for example a common next word, or a word you often say next.
The prediction runs entirely on the device, with no network calls. It is a trigram model built from 666 sentences written for Voice, with an optional personal layer and a small ranker on top. Each set of suggestions takes well under a millisecond on a desktop computer.
Learning is off by default. When it is switched on, Voice learns only from messages that are actually spoken, never from taps that are deleted. Here, after I want Grandma has been spoken three times, Grandma appears as a suggestion, marked as a word you often say next. Everything learned can be seen, and forgotten, one word at a time.
A caregiver dashboard sits alongside the app. You can try the board, the full app and the dashboard, all with made-up data, on Paul's portfolio at paulmartinmcneill.com.
-
Flood warnings without the weather forecast
Transcript
Can river floods be forecast without knowing tomorrow's weather? Paul McNeill tested that across 671 British catchments, using CAMELS-GB, a public dataset of daily river flow and weather from 1970 to 2022.
The task: on any given day, forecast river flow one, three and seven days ahead, using only what is known that day. The models were trained on 1970 to 2004, tuned on 2004 to 2012, and tested on 2012 to 2022.
Skill is measured with a score where 1 is a perfect forecast and 0 is no better than the long-term average. Most of the skill beyond a day depends on knowing the weather. At seven days, the median score is 0.24 without future weather, and 0.72 when the model is given the weather that actually happened. That second number is an upper bound, not a forecast.
Forecasting a whole region the model never saw costs little: a median loss of 0.019 at one day. A different climate costs much more. Trained on wet catchments and tested on dry ones, the score falls from 0.85 to 0.75.
A neural network, an LSTM reading a year of daily data, did no better than gradient boosting. That points to the missing weather forecast as the limit, not the choice of model.
Warnings are harder still. At seven days, most high-flow warnings are false alarms. The study has not been peer reviewed, and every decision, including the mistakes, is logged in public.
-
Why most cancer drug response stays unexplained
Transcript
One gene, PTEN, is a known but incomplete guide to how cancer cells respond to drugs that block a protein called AKT. This study asks how much more wider molecular data can explain, using 924 cancer cell lines from the public GDSC screen.
PTEN alone explains about 5 per cent of the variation in response. Adding cancer type and sequencing centre takes that to 11 per cent. Adding two thousand genes reaches 14 per cent with an elastic net, and 19 per cent with gradient boosting.
So more than four fifths of the response stays unexplained. And that gap is not mainly measurement noise. A related drug was measured twice: about 80 per cent of its variation repeats, yet the best model for it explains only 18 per cent.
The study also measures a common mistake. Choosing genes by their link to the response before cross-validating produces a score of 0.14 even on shuffled data, where there is nothing to find. Choosing inside each fold removes the leak.
The explanations were not stable either. The top twenty genes of refitted models overlap by only 0.2, and no stable subgroups were found beyond tissue of origin.
Cell lines are a stand-in: nothing here is about patients or treatment. The study has not been peer reviewed. The code, protocol and decision log are public.
-
Six rules for research you can check
Transcript
Two studies, one set of rules. This is how Paul McNeill runs machine learning research that someone else can check.
Protocol first: the design is written and dated before any model is fitted, and any later change is a dated amendment.
Baselines first: every model is compared with the simplest thing that could work, on exactly the same held-out data. The data splits are saved once with a fingerprint, and the code refuses to overwrite them.
Nothing is learned from held-out data: feature choice, scaling, gap filling and tuning all happen inside the training data. What counts as this did not work is set in advance, and negative results are reported.
And every decision is logged, including the mistakes. In the river study, the first diagnosis of a weak baseline was wrong; a controlled comparison showed the cause was the target scale. In the drug study, gains were first computed on averaged predictions, which flattered the less stable model. Both were corrected in public.
Both studies were built making use of AI tools, under research rules Paul set, and each study's log records how. Re-run on a second machine, the main results matched almost exactly; one river experiment is still to be repeated. Neither study has been peer reviewed.
-
A portfolio you can listen to and talk to
Transcript
Paul McNeill's portfolio is built to be heard as well as read. The main pages have a short recorded narration, with the full transcript one tap away.
The narration uses an AI voice made with ElevenLabs, and every player says so. It is not Paul's voice.
The site's text assistant can be used by voice too. Press the microphone and speak; the question appears as you talk, and the reply can be read aloud.
Visitors can choose that voice, the same way they choose the colour and the font. The picker lists the voices on their own device, with a speed control and a preview, and the choice is remembered in their browser.
On the Lab page there is also a live voice assistant. It answers questions about Paul's work from a fact sheet taken from the site, and says it is not Paul. Nothing from ElevenLabs loads until the visitor presses the button.
All of it is plain HTML, CSS and JavaScript on a static site. Speech recognition and the voice picker use speech features built into the visitor's browser, and the site still works with no sound at all.
-
Four AI demos in one web page
Transcript
The Lab on Paul McNeill's portfolio holds four working demos. Three of them run entirely on the visitor's device.
First, a sentiment model. It is a small public example model from TensorFlow.js, trained on movie reviews. Type a sentence and it scores how positive it is, inside the visitor's browser, with nothing they type sent to a server.
Second, the word board from Voice, Paul's AAC app. It runs the app's real prediction engine, so the suggestions and their reasons match the app.
Third, a live voice assistant built with ElevenLabs. This is the one demo that uses a server: what the visitor says is sent to ElevenLabs, and the page explains that before anything loads.
Fourth, an animation of how a research agent works through a task: input, planning, tool use, memory, synthesis and output. It is an illustration of the pattern, not a live agent.
Narrated with an AI voice made with ElevenLabs. It is not Paul’s voice. Captions are part of each video, and every one has a transcript.
What went wrong, and what changed
The early versions chased engagement: daily news topics, fast cuts, stock imagery. The results were busy and hard to verify. The explainer series reverses that: fewer, slower videos about work I can stand behind, where every claim has a source.
Tech stack
Python, Anthropic API (Claude), ElevenLabs, Edge TTS, Whisper, Playwright, Pillow, FFmpeg and libass, PyYAML, GitPython.