/ˈskep.sis/ · noun, Ancient Greek
Inquiry; the act of thinking, examining, and weighing before concluding. The discipline of thought itself.
Skepsis Labs builds reasoning systems on open-weight language models — search that understands meaning, agents that draft and examine long documents, and the inference infrastructure to run frontier models on hardware we own.
Shoppers describe what they want; keyword search returns what they typed. We are building embedding-based semantic product discovery for e-commerce platforms — search that matches intent, not strings — starting with an app for Shopify merchants.
Retrieval is served from our own GPU cluster, keeping per-query economics viable at long-tail merchant scale.
Technical grant applications are long-context reasoning problems: a funder's scheme documents, a company's technical evidence, and a persuasive narrative that must reconcile both. We are building agentic drafting and analysis workflows for structured technical documents, beginning with UK innovation-funding applications for deep-tech SMEs.
Open-weight mixture-of-experts models now reach frontier capability at hundreds of billions of parameters. We research and engineer the serving side: multi-GPU tensor-parallel deployment, hybrid GPU–CPU expert offload, quantization trade-offs, and long-context prefill optimization — so that trillion-parameter-class models run privately, on-premises, on accessible hardware.
Everything we ship is developed and served on NVIDIA accelerated computing that we operate ourselves. Owning the inference layer is a product decision: it sets our cost floor, keeps customer data on hardware we control, and lets us work at the quantization and serving frontier rather than behind an API.
Skepsis Labs is an independent AI lab founded in 2026 by Alexandros Triantafyllidis — a machine-learning engineer with twenty years in natural language processing, from statistical NLP through the transformer era to frontier large-language-model work; MSc in NLP, University of Edinburgh.
The lab designs, evaluates, and serves open-weight models on its own NVIDIA infrastructure, end to end.
Working on search, documents, or local inference — or want to compare notes on serving open-weight models?