AksharaCheck · अक्षर जाँच · ಅಕ್ಷರ ಪರಿಶೀಲನೆ

Find the letters your website breaks. AksharaCheck checks Hindi and Kannada text in Chromium, records visual and Unicode evidence, and optionally uses BharatCode to explain actionable findings and propose repairs. A proposed repair is marked verified only when a fresh Chromium check passes.

Recorded result: 9 of 9 actionable findings explained by BharatCode · 4 repairs verified in Chromium · public-page examples in Hindi and Kannada.

Website and text checker · BharatCode demo (verified repairs) · Kannada public-page report · Hindi public-page report · BharatCode.txt

A three-minute tour for judges

1. See real BharatCode output. Open the controlled demonstration report, switch between English, हिंदी and ಕನ್ನಡ, and inspect the overflow finding's rejected and accepted repair attempts. The report contains the actual before/after Chromium screenshots.
2. See a real Kannada finding. Open the Karnataka Varthe report: 72 pieces of Kannada text on a government news site are tagged as English because the page's <html> element says lang="en". The report traces all 72 to that one tag.
3. See a clean Hindi scan. The DD News report shows what a pass looks like: what was checked, how many text elements, and both page screenshots. Then try the in-browser text checker.

The public GitHub Pages site is a static showcase and text checker, not a live URL scanner. Run the Python/Chromium scanner locally using the instructions below.

The problem

A page can look fine during a quick visual review but contain malformed Indic character sequences, clipped marks, phone-width overflow or incorrect language metadata. For example, a Kannada passage tagged as English may be pronounced incorrectly by assistive technology. AksharaCheck records browser evidence so developers can investigate and fix these issues without relying on model guesses.

How it works

flowchart LR
    A[Chromium opens the page<br/>at 390 px and 1440 px] --> B[Rules find evidence:<br/>letters, fonts, clipping,<br/>overflow, language tags]
    B --> C[Optional BharatCode AI for<br/>actionable findings: explanations<br/>and proposed CSS or lang repair]
    C --> D{Fresh Chromium<br/>tests the fix}
    D -- passes --> E[Verified fix<br/>with before and after]
    D -- rejected, with the reason --> C

Loading

1. Chromium measures. At phone (390 px) and desktop (1440 px) widths, the scanner records sampled text, character sequences, language tags, font evidence, geometry and screenshots. Browser rules produce the findings; the model does not invent them.
2. BharatCode explains (optional). With --ai, BharatCode's qwen-3.8-27b receives actionable Fix/Check findings, explains them in English, provides Hindi/Kannada summaries and may propose a restricted CSS or lang repair. Informational notes remain browser-only.
3. Chromium verifies. Proposed repairs are tested on a fresh page. A repair earns verified status only if the original issue clears and visibility/size checks pass. Rejected attempts and reasons remain in the report, with at most three tries. The scanner never changes the target website.

Results so far

Recorded scan	What it shows
BharatCode demo report (a test page with planted bugs, not a real website)	3 problems to fix, 6 to check, 2 notes. BharatCode explained 9 of 9 findings in three languages; Chromium verified 4 repairs, including one that passed only after a first suggestion was rejected.
Karnataka Varthe (Kannada), 31 Oct 2026	72 pieces of Kannada text tagged as English, all inherited from the page's <html lang="en">, so one change fixes them. One possible phone overflow to review and 13 font notes. No problems in the Fix category.
DD News (Hindi), 31 Oct 2026	No findings at either width. Chromium checked 162 text elements on the phone layout and 213 on desktop, and the report shows both page screenshots.

These are dated snapshots of one page at two widths, published exactly as AksharaCheck wrote them. The demo shows BharatCode's explanations and verified repairs; the public-site reports are browser checks only.

Fewer false alarms is part of the result. Earlier DD News scans produced 63 notes about line spacing and fonts. Narrowing those rules to skip ordinary Hindi line spacing and Windows' expected Arial-to-Nirmala UI font substitution brought it to zero, a change in the scanner rather than in DD News.

What it checks

Check	Level	What it means
Broken letters	Fix	A vowel sign, nukta or virama with no letter to attach to
Web font did not load	Fix	A font file the page asked for failed to download
Wrong language tag	Check	Text in Devanagari or Kannada script tagged as another language
Two languages without labels	Check	Hindi and Kannada in one element under one language tag
Old non-Unicode font	Check	Latin letters shown through a legacy font such as KrutiDev
Text runs off the screen	Check	A line wider than its box at phone width
Letters may be cut off	Check	Text taller than the box that shows it
Font family names differ	Note	The CSS family name differs from Chromium's reported internal font name; this can be intentional
Unusually tight line spacing	Note	Multiple text lines use unusually little vertical separation; review visually

Run a browser-only scan (no BharatCode key needed)

You need Python 3.10 or newer and permission to scan the selected public page. On Windows Command Prompt:

git clone https://github.com/ashh-droid/AksharaCheck.git
cd AksharaCheck
py -3 -m venv .venv
.venv\Scripts\python.exe -m pip install -e ".[test]"
.venv\Scripts\python.exe -m playwright install chromium
.venv\Scripts\python.exe -m aksharacheck scan https://hi.wikipedia.org/ --lang both --output output\first-scan
start output\first-scan\report.html

On macOS or Linux, use python3 -m venv .venv, .venv/bin/python in place of .venv\Scripts\python.exe, and open or xdg-open for the report.

The command above runs the browser checks and writes report.html, report.md and report.json. No account or API key is needed. Optionally add --ai to request BharatCode explanations and repair verification for actionable findings; this requires a valid BHARATCODE_API_KEY and may incur model usage. Use --fail-on error in CI if you want a nonzero exit code when a Fix-level finding is recorded.

Reproduce the BharatCode demonstration (optional)

For judges: open the published 9/9 demo report. It already includes genuine BharatCode outputs, four Chromium-verified repairs and rejected-attempt evidence; you do not need an API key to view it.

To run the model-assisted fixture again yourself (requires a BharatCode key and new API calls after detector changes):

.venv\Scripts\python.exe scripts\run_ai_demo.py --max-findings 9
start output\ai-demo\report.html

The script serves fixtures/toota-akshar, scans it and runs the BharatCode/Chromium loop. The key is entered at a hidden prompt and is not saved. Successful findings can be checkpointed and resumed only when the scanner, AI client and fixture have not changed; older checkpoints are invalidated after relevant code changes. See development instructions for troubleshooting.

Run the local web app

Start the Python API in one terminal, from the repository root:

.venv\Scripts\python.exe -m aksharacheck serve --port 8000

In a second terminal, from the same repository root:

cd web
npm install
npm run dev

Open http://127.0.0.1:5173. The local interface can request scans through the running Python server; GitHub Pages hosts only the static showcase and in-browser text checker.

Project structure

Path	Contents
aksharacheck/browser.py	Playwright scanner: phone and desktop widths, fonts, screenshots
aksharacheck/detectors.py	The rendering rules
aksharacheck/ai.py	BharatCode client, CSS whitelist and the Chromium verification loop
aksharacheck/report.py	HTML, Markdown and JSON reports
aksharacheck/api.py, cli.py	FastAPI server and command line
web/	React and Vite website, including the in-browser text checker
scripts/	BharatCode demo and key check, report publishing, visual QA and Windows setup
fixtures/toota-akshar/	Test page with planted rendering bugs
tests/	Automated unit and real-browser checks of the website and reports

Testing

.venv\Scripts\python.exe -m pytest -q

GitHub Actions runs the full suite on every push, including Chromium checks of the website at nine screen widths and of a generated report at five. scripts\check_report_ui.py <report.html> checks any report the same way.

Security and privacy

• The BharatCode key stays in a server-side environment variable. It is never sent to the browser, written to the checkpoint or printed.
• Page content is treated as untrusted. The model can only return whitelisted CSS properties and a lang value; URLs, imports and selectors are rejected.
• The scanner refuses local and private network addresses unless --allow-local is passed on the command line for a trusted test page.
• Reports include text samples from the scanned page, so scan public pages only.

Limitations

AksharaCheck samples visible page text in Chromium at the recorded widths and captures; it is not a complete WCAG accessibility audit or a guarantee about every page, device, browser or screen reader. It does not review PDFs, images, native apps, login-only pages, spelling or translation. Check/Note findings need human review of the underlying element and screenshot. Repairs are tested on temporary browser copies only, never applied to the scanned website.

Development disclosure

AI tools including ChatGPT and Claude assisted with planning, coding, debugging, design and documentation. The team reviewed and tested the work; the test suite and the Chromium checks are reproducible. The nine-finding demonstration used real BharatCode API responses and fresh browser checks, and the report marks which explanations were reused from an earlier run.

The report fonts are Anek Latin, Anek Devanagari and Anek Kannada, under the SIL Open Font License (aksharacheck/assets/OFL.txt).