தமிழ்AI
International Educational Foundation

தமிழ்AI

Two thousand years of Tamil grammar, in a form a machine can cite.

An AI can write fluent Tamil today. Ask it how a Tamil word is built, or which rule makes that true, and it is often wrong, and confident about it.

தமிழ்AI is the part underneath that does not have to guess. A free, open grammar engine that takes a Tamil word apart, names the நூற்பா behind every claim, and returns an honest gap when no source can settle the answer.

How it is built Why it is needed Code on GitHub

வட்டெழுத்து inscription, Brihadeeswara temple, Thanjavur · CC0 · Wikimedia Commons

9
tools live
266
tests passing
94/108
origin correct
11
honest unknowns

Measured on a 108 word everyday sweep. One answer is wrong and we name it. Verified 2026-08-11 against CODE-STATUS.md.

What it is made of

Four layers, and a horizon we are honest about

Each layer only works because of the one above it, and the top layer feeds data back down. Open a layer to see what its page carries, or ask for the fuller explanation.

Layer 1

இலக்கணம் grammar

What makes an answer true, and who says so.

What the page carries
  • Which authority governs which topic, as a table
  • How a citation is written: அதிகாரம் › இயல் › நூற்பா
  • Why we pinned full editions instead of trusting quoted verse numbers
Layer 2

பகுபத உறுப்பிலக்கணம் morphology

How a Tamil word is taken apart, and why a dictionary can never do it.

What the page carries
  • Agglutination explained without jargon
  • The six உறுப்பு, each with its technical name
  • One word shown as anatomy, each part labelled with its role and its authority
  • ThamizhiMorph credited in full, and the wider Thamizhi suite mapped onto our roadmap
Layer 3

மூலம் sources

What it grounds on, what we still need, what we may not use.

What the page carries
  • The ledger: every source, its evidential grade, and what we may legally do with it
  • A page each for ThamizhiMorph, Project Madurai, TVA, Aalamaram, ILAKKANAM and the Madras Lexicon
  • Three open asks, with what we want and what we would give back
  • The one licence gap we currently ship, named in public
Layer 4

பகிர்வு distribution

How this reaches a Tamil speaker who will never install anything.

What the page carries
  • The nine tools, each with its objective in one line, and the chain that connects them
  • What each tool refuses to do, which is the part that earns trust
  • Where the server will be installable from, and which registries list it
  • What ships as data, and under what terms
Horizon

இயல்பு தமிழ் படிமம் a Tamil-native model

Written down, deliberately not scheduled.

What the page carries
  • The stages from word analyser to sentence, in order, each shippable on its own
  • The data flywheel, drawn
  • The two questions we have to settle before any of it starts

every use feeds verified data back down ↑

Layer 1

இலக்கணம் grammar

Ask a model a Tamil grammar question and it answers from whatever it read on the internet. That is a popularity contest. We pick the authority first, then make the software obey it.

தொல்காப்பியம் is first. நன்னூல் is used where Tholkappiyam does not codify the point, and the six-part பகுபத உறுப்பு scheme is the main such place. Every answer records which of the two it used.

Both editions are pinned as version-locked, checksummed files and read while the server runs. So when the tool says a word is in the seventh வேற்றுமை, it shows the நூற்பா that says so. You get to disagree with the verse rather than with a machine.

One trap this already cost us: verse numbers quoted in secondary sources drift. Three of our rule tables carried wrong numbers before we pinned the full texts.

Read the full page →

Layer 2

பகுபத உறுப்பிலக்கணம் morphology

Morphology is the study of how a word is built out of parts. Tamil grammar has had a name for that for centuries. Same subject, older vocabulary.

Tamil keeps its grammar inside the word. English needs three words for "in the tree", Tamil needs one. A system that treats a Tamil word as a single opaque string has thrown away most of the grammar before it starts.

And the forms cannot be listed. One verb root yields thousands, so a fixed word list is out of date the moment somebody inflects something. That is why the core is a rule engine and not a database.

The rules run on ThamizhiMorph, the finite-state analyser published by Sarveswaran, Dias and Butt in 2021 under Apache-2.0. A finite-state machine runs in both directions: it can take a word apart, and it can build one. We use one direction today. The other is where the long road starts.

Read the full page →

Layer 3

மூலம் sources

Two questions get asked about every source, and they are kept apart on purpose. How good is the evidence, and what are we allowed to do with it. Collapsing them causes mistakes in both directions.

So each source carries a grade from A to D, and separately a redistribution mode: ship it, serve it with attribution, or consult and cite it only. The Madras University Lexicon is excellent evidence and we may not bundle a line of it. Project Madurai’s classical texts travel with us, because they grant distribution as long as the header stays intact.

Held today: the pinned Tholkappiyam and Nannūl, ThamizhiMorph, Tamil Wiktionary, and our own verified verb paradigms.

Asked for: the TVA கலைச்சொல் glossaries, the Aalamaram treebank, the ILAKKANAM benchmark. Each has a page saying what we want and what we would give back.

Read the full page →

Layer 4

பகிர்வு distribution

An MCP server is not something a person uses. It is an adapter that lets an AI assistant call your logic. So the thing to build is one engine with several front doors.

Today there are four: an MCP head for assistants, a REST API for apps, a browser page for people who will never install anything, and a command line. The browser page needed no engine changes at all, which is how we know the design held.

Nobody running it locally is required to set up a container or a database. That zero-config install is protected on purpose, because the people most likely to try this are teachers, not sysadmins.

The code is Apache-2.0 under a nonprofit and adoption is the goal. Anyone can self-host, and the curated data goes to Hugging Face where the people training models actually look.

Read the full page →

Horizon

இயல்பு தமிழ் படிமம் a Tamil-native model

Tamil is short of machine-readable linguistic data, not of literature. So every analysis the server resolves is stored as verified, provenance-tagged data. The corpus grows as a by-product of ordinary use rather than as a project somebody has to fund.

Enough of it makes two things possible. A tokenizer whose units are morphemes instead of statistical fragments, which attacks the token explosion problem at its root. Then a small Tamil model trained on data that was checked rather than scraped.

None of this is scheduled, and the page says so. There is a written trigger: roughly 100,000 verified records and evaluation that runs as routine. Until both hold, the work stays where it is.

The endpoint we expect is a pairing. A large model handles meaning, this engine handles morphology and word building. Cheaper to run than either half pretending to do both.

Read the full page →

இலக்கணம் alone

A book. It knows the rules, and it cannot apply them to the word you just typed.

Morphology alone

A machine. It splits the word, and it cannot tell you which authority says the split is right.

Put them together and the answer is computed and citable at the same time. That is the whole product, and everything else on this site is detail.

The site

Sixteen pages, each answering one question

Ten core pages carry the argument. Six more give each source and institution a page of its own, so a letter can link one address instead of explaining itself.

Core

/

What is this, and what is it made of?

/why

Why does a language this well documented count as low-resource?

/grammar

What makes an answer true here?

/morphology

How is a Tamil word taken apart?

/engine

What have we actually built? The nine tools and how they chain.

/sources

What do we ground on, and what may we use it for?

/distribution

How does it reach people?

/horizon

Where does this go, and what is not scheduled?

/status

What works today, measured?

/glossary

What does this term mean in Tamil grammar?

/about

Who is doing this, and how do I write to them?

One page per source

/sources/thamizhimorph

The FST we wrap, and the wider Thamizhi suite. The build starts here.

/sources/project-madurai

The pinned Tholkappiyam and Nannūl every claim quotes.

/sources/tva

The கலைச்சொல் glossaries we are asking for.

/sources/aalamaram

The treebank we have adopted and not yet obtained.

/sources/ilakkanam

The benchmark we measure ourselves against.

/sources/madras-lexicon

Closed with a reason, published rather than hidden.