We build AI products on frontier models, train custom models where APIs fall short, and take both to production — while publishing research that advances the field.
100+ projects for enterprise and growth-stage companies. We build AI products on frontier APIs and on-prem open models, train and fine-tune custom models, and run the discovery work that finds the highest-value use cases.
Anyone can build on an API. We train, pretrain, and fine-tune models — and publish the results. Current work: Physical AI — world models, robotic failure detection, and benchmark science.
From custom model training to full production deployments — across every AI modality.
We train JEPA-based world models with long-horizon supervision, build and benchmark robotic failure detection (FailBench, FailDetect), and run meta-analysis across the Physical AI benchmark landscape (MetaBench Physical). Papers are published or under review; models and benchmarks are headed to open source.
The same expertise ships in client work: anomaly detection, factory monitoring, defect detection, and traffic intelligence — deployed at production scale.
See our Physical AI research →Improving intuitive physics understanding through long-horizon supervision, plus action-conditioned JEPA world models steered by image goals and text instructions.
A pooled benchmark for robotic arm failure detection — evaluating specialist models against generalist VLMs — and FailDetect, a model that zooms, crops, and catches failures.
Every vendor reports on different benchmarks. MetaBench Physical builds one comparable matrix, prunes redundant benchmarks to an optimal subset, and ranks models on it.
Also from the lab: low-resource text embeddings, published at EACL 2026 · #1 held on ViDoRe visual document retrieval · Uzbek legal embeddings & RAG, from a client engagement.
Full voice pipeline — ASR, LLM, TTS — with fine-tuned models for Armenian, English and Russian. Sub-300ms latency, 98.7% accuracy.
Industrial anomaly detection deployed across 12 production lines. 40% reduction in defect pass-through.
Multimodal RAG pipeline over complex legal documents. Fine-tuned for jurisdiction-specific reasoning.
We respond to every serious inquiry — usually within a day.
We'll be in touch soon.