antiQ.ai

antiQ.ai is building an AI‑native ecosystem for ancient language understanding and cultural analysis. Latin is the current sandbox (has the most demand across dead languages), not our ceiling: by solving morphology, syntax, and translation at scale we create reusable infrastructure for every "dead" language that still shapes the living world.

Live features

• Variant‑text reader — original, literal, and idiomatic translations aligned line‑for‑line. Also offers a Simplified Latin layer—useful now, but likely redundant once the texts are solved.
• One‑click lookup + flashcards — tap a word, see the dictionary entry, and drop it straight into an SRS deck that adapts as you read.

Next up

Early users are asking for guidance, instead of spoilers. A full translation—or even a pre‑solved dictionary form—feels like “cheating.”

To that end, Q2 2025 will launch the Smart Reader (will find a better name) that knows what you know and don't, and reveals only the minimum hint required to keep you moving. This will tie in nicely with Flashcards (vocabulary-learning) & Gym (grammar-learning -- though this is not operational right now).

This requires getting definitive morphology and syntax tags for each text. This, it turns out, is quite hard. I am trying to progress in 2 phases (morphology, and syntax) so to solve word-level and phrase-level issues.

Enabling Engines

• Morphology model ✓ (not functional). Probably SOTA accuracy for Classical Latin. Once a pipeline, word-lookup will be near flawless and simulate an interaction with a tutor when asked for a what word is (What is the ending? Why does it end with -is? What does aliam do in that sentence?). A "workflow" based on this model cuts the error in half.
• Syntax model (Q2 2025). The basic idea is to be able to detect troublesome syntactical constructions that are not straightforward to resolve without a grammar book or commentary. Thinking of using Dependency Graphs for this.

Bigger picture

Once morphology and syntax are machine‑solved, the Latin corpus becomes a fully structured dataset (practically a solved language). Pushed into a graph database, that structure lets us perform a Foucault‑style archaeology of ideas—querying how concepts mutate from Cicero through Augustine to Erasmus. The same pipeline ports to Greek, Sanskrit, or any archive waiting for its second life.

Benchmark snapshot for morphological tagging (Apr 2025 - for Classical Latin (Perseus Treebank))

Recipe: Self-training with high confidence sentences, targeted synthetic data generation, and more frozen pre-training on classical corpus. The remaining errors might have to do with annotation disagreements rather than with the architecture and training. In practice, of course, these will be solved by general purpose LLMs or crowdsourcing.

Status — Public Beta

Beta is open to everyone. Try it at antiq.ai and help shape the roadmap.