Kalvettu (கல்வெட்டு): read a Tamil-Brahmi inscription from a photo
Photograph a 2,000-year-old Tamil inscription on a cave, a rock or a potsherd, draw a box around
each letter, and Kalvettu tells you what each letter is, how sure it is, and every way the line
can be read. It never makes up a meaning.
Tamil-Brahmi is the script of the earliest Tamil inscriptions, from about the 3rd century BCE to
the 4th century CE. It is carved on the rock shelters of Tamil Nadu's hills (Mangulam, Jambai,
Arittapatti and many more), but almost no visitor can read it, and the published readings sit in
books and journals. Kalvettu is a first step: a letter reader that runs on a phone, offline, and
teaches the script along the way.
Open it: https://havenleno22.github.io/kalvettu/ (works in any phone or desktop browser; add it
to the home screen to keep it offline).
What you can do
| Page | What it is for |
|---|
| Read | Take or choose a photo, drag a box around each letter (move it, resize it, or Tighten box if it is loose), and see each letter with how sure the reader is, its other possibilities, and the whole line in Tamil-Brahmi, Latin letters (ISO 15919) and Tamil |
| Learn | Five short lessons (how the script works, the vowels, the eighteen consonants, the vowel signs, the dot and the three spellings), then practice: a letter on the stone, four answers |
| Write | Type a word or a name in Tamil (or tap the letters) and see it in Tamil-Brahmi, spelled the way each period would have spelled it; save it as a picture or copy it as Unicode text |
| Guide | Step by step from finding an inscription to reading it: photographing, drawing boxes, what the marks mean, a glossary and common questions |
| Letters | Every letter and vowel sign, and one word in three spellings |
| About | How the reader was made and measured |
The promise: nothing made up
- It reports what it sees. Every letter comes with how sure the reader is. Under 60% it says
"not sure" and shows the possibilities instead of guessing.
- It shows every reading. Tamil-Brahmi was spelled three ways over the centuries (Iravatham
Mahadevan's TB-I, TB-II and TB-III), so the same letters can sound different. Kalvettu shows each
reading instead of quietly picking one.
- It never makes up a meaning. A meaning will appear only when a published scholar's reading
matches, with the source named. That part is not built yet.
How accurate is it?
Each number with what it was measured on (details in docs/model.md):
| Test | Result |
|---|
| Held-out synthetic validation set, 20,000 crops (base letter, vowel sign and dot all right) | 86.6% |
| Calibration error on the same set (how far "90% sure" is from being right 90% of the time) | 1.8% |
| A real photograph of the Jambai inscription, 21 letters | 18 of 21, draft |
The Jambai result is not yet a fair test: the labels have not been checked by an epigraphist, and
the same inscription was used while building the training data. A clean test set from other
inscriptions, labelled by an expert, is the next step.
How it works
photo ─► your boxes ─► tighten to the strokes ─► 96×96 crop ─► letter reader (ONNX, in the browser)
│ base letter, vowel sign, dot
▼
readings under TB-I, TB-II, TB-III ◄── calibrated probabilities, impossible forms removed
- Letter reader. A small residual CNN (2.8 M parameters, 11 MB) with three heads: the base
letter (34, plus "not a letter"), the vowel sign, and the dot (puḷḷi). It runs in the browser
with ONNX Runtime Web, so photos never leave the device.
- Training data. 1.2 million synthetic letters: shapes from two open-licence fonts, recombined
into forms the fonts do not draw, redrawn stroke by stroke with random pens and bends, then
rendered as rock carving in side light, chalk-filled carving, estampage, potsherd scratching or
ink drawing, and put through a simulated camera.
- Reading rules. The three writing systems, written once in Python and once in TypeScript and
tested against the same hand-written cases (spec/reading_cases.json).
- Script reference. Everything the code assumes about the script is written down with its
source in docs/tamil-brahmi.md, following Unicode 14's Old Tamil
additions and Mahadevan's Early Tamil Epigraphy.
Run it yourself
The app (Node 20 or newer):
cd app
npm install
npm test # reading rules, letter logic, box tightening, Tamil to Tamil-Brahmi
npm run dev # http://localhost:5173
The machine-learning side (Python 3.11 with uv) needs two fonts that
are not in this repository. Put them in references/:
references/adinatha/Adinatha-Tamil-Brahmi/Adinatha-Tamil-Brahmi.otf: Adinatha Tamil Brahmi
(Vinodh Rajan, Shriramana Sharma and Udhaya Sankar, 2012, SIL Open Font License), from
virtualvinodh.comreferences/NotoSansBrahmi-Regular.ttf: Noto Sans Brahmi (SIL Open Font License), the same file
as app/public/fonts/NotoSansBrahmi-Regular.ttf
cd ml
uv sync
uv run pytest -q
docs/PROGRESS.md has every command for making data, training, evaluating and
exporting a model, and what comes next.
Layout
| Folder | What is in it |
|---|
app/ | The web app (React, TypeScript, Vite, ONNX Runtime Web), offline-first |
ml/ | Python: sign inventory, reading rules, synthetic data, training, evaluation, export |
spec/ | Files both halves share: the inventory and the reading test cases |
data/ | Licence records and draft labels; images and generated data stay out of the repository (data/README.md) |
docs/ | Design, script reference, model card, progress |
Licences
- Code and the trained reader: MIT, see LICENSE.
- Letter outlines in
app/public/glyphs/ are derived from Adinatha Tamil Brahmi and keep its SIL
Open Font License (text); they do not use its
reserved font name.
- Noto Sans Brahmi in
app/public/fonts/: SIL Open Font License (text).
- The shipped reader was trained only on synthetic images. Photographs from Wikimedia Commons are
used for testing under their own licences, recorded in
data/real/*/manifest.jsonl.
Help wanted
If you read Tamil-Brahmi professionally, switch on Expert labels in Read, correct the letters
and export them. Checked labels are what Kalvettu needs most. Open an issue, or write to
[email protected].
Built by Haven Leno J ·
LinkedIn · Portfolio
