Three active research directions: Physical AI (primary focus), Armenian & low-resource language AI, and visual document understanding. We publish openly and put models on Hugging Face.
We believe the most interesting unsolved problems in AI are about understanding and acting in physical environments. Our research sits at the intersection of two hard problems.
Training VLMs to reason about spatial structure, scene geometry, and physical constraints — moving beyond "what is this" to "where is this, how does it move, what will happen next."
Joint Embedding Predictive Architectures (JEPA) for long-horizon robotic planning. Teaching models to predict useful representations of future states, not just next tokens.
Submitted to NeurIPS 2026. JEPA-based architectures trained for extended planning over physical action sequences in robotics tasks.
We are Armenians. We invest our own budget and time into building AI infrastructure for Armenian — not because it's a business, but because it matters. The methods we develop translate directly to Georgian, Uzbek, and any other low-resource language.
Armenian Text Embeddings — SOTA embedding models for Armenian. Outperform Gemini and OpenAI embeddings on every Armenian benchmark.
#1 on ArmBench-TextEmbedArmBench-TextEmbed and ArmBench-LLM — the first open benchmarks for evaluating text embedding and LLM performance in Armenian.
First Armenian benchmarksTraining an LLM with a localized tokenizer from scratch. Existing models are tokenizer-inefficient for Armenian — we're fixing that with our own data pipeline.
Active trainingSOTA embedding quality in a low-resource language using only 10k noisy translations — no large parallel corpora required. Applied to Armenian; validated on Georgian and Uzbek. Method generalises to any low-resource language.
We built ColPali-style multimodal embeddings for visual document retrieval and held the #1 rank on the ViDoRe benchmark globally. The field matured and became crowded — we turned our attention to harder problems.
Visual document retrieval embeddings ranked #1 on ViDoRe benchmark. Models still available and used in production. Research direction discontinued.
All our models and benchmarks are public. We believe open science makes everyone better.