Project Progress
Where the project stands, what shipped recently, and what is still being decided. The source of truth is the milestones on GitHub. This page is a summary of them.
At a Glance
- PowerPoint is working end to end. Titles go first in the reading order, missing alt text is written, missing language is detected, and small text is reported. Both the website and the folder script work.
- PDFs are now accepted. They are routed page by page and rebuilt as tagged PDFs. Scanned pages use OCR and always go to a person for review.
- Not deployed yet, and there is no login. Visitors are kept apart only by a signed browser cookie. University sign-in is planned.
- The alt-text model is still undecided. The main branch runs a local model (about 4.4 GB). A branch that switches to Claude on Amazon Bedrock is waiting for a decision.
Milestones
Issues closed out of the total in each milestone, ordered by target date.
| Milestone | Done | Target |
|---|---|---|
| Faculty web intake | 1 of 1 Complete | |
| Slides: core checks | 5 of 9 | |
| Reports and coverage docs | 1 of 4 | |
| Documentation and accessibility audit | 0 of 2 | |
| Slides: alt text and media | 0 of 9 | |
| Accessibility conformance | 0 of 8 | |
| Observability and fault tolerance | 0 of 1 | |
| Deployment and access control | 0 of 8 | |
| Orchestration and notifications | 0 of 2 | |
| Submission dashboard and reporting | 0 of 2 | |
| Beyond slides: web pages and other formats | 2 of 16 | |
| Script and code health | 0 of 9 | No date |
| Future fixes | 0 of 1 | No date |
Recently Shipped
- PDF remediation (). Uploaded PDFs go through a page-by-page pipeline and come back as tagged PDFs, with OCR for scanned pages and human review for anything uncertain. #84, #92
- Presentation language (). Missing, invalid or mixed language settings are reported, and can be set across the whole deck. #24
- Minimum font size (). Text below 9 pt is reported by slide, including sizes inherited from layouts and masters. #26
- Title detection by layout (). When a slide has no real title placeholder, a small trained model picks the likely title from size, position and wording. #7, #23
- Documentation (). A WCAG 2.1 coverage table, notes on exporting fixed decks to PDF, and a decision on Prezi (export to PowerPoint first). #50, #37
Open Decisions
- Local model or hosted API for alt text. A local model needs a GPU machine and about 4.4 GB of files. A hosted API runs on a lighter server but costs money per image. #97
- Whether n8n should orchestrate the pipeline, which affects how notifications and storage are built. #80
- How to check web pages for problems that only show once rendered, such as CSS contrast and keyboard access. #47
Early Results
Faculty PowerPoint files that Anthology Ally flagged for missing headings were run through the tool and put back into their Canvas courses. A full comparison with other checkers is tracked in #98.
| File | Ally score | Flagged issues |
|---|---|---|
| A Beginner's Guide to SEO | 51% to 99% | 48 to 3 |
| Ultimate Guide in Slides (Dec 2021) | 55% to 68% | 69 to 4 |
| Mini Media Plan Presentation Template | 48% to 68% | 2 to 1 |
| Content Marketing | 49% to 61% | 55 to 20 |
| Ultimate Guide in Slides (Jun 2020) | 28% to 48% | 76 to not recorded |
Every score went up. The issues left on the SEO lesson are all color contrast, which the tool reports but does not fix yet.