Former DeepMind researchers at Poetiq secure $45.8M to enhance open source and proprietary LLMs through a meta-system that improves accuracy and slashes compute costs, positioning software optimisation as a faster, cheaper path than retraining.
Poetiq Inc. has raised $45.8 million in seed funding to scale a software layer designed to enhance large language models, including open-source systems, without rebuilding or retraining them. The round was led by FYRFLY Venture Partners and Surface Ventures, with participation from Y Combinator, 468 Capital, Operator Collective, Hico Ventures and Neuron Venture Partners, signalling early investor confidence in tools that optimise existing models rather than create new ones.
The startup has developed a “meta-system” that sits on top of LLMs such as Google’s Gemini series, GPT-5.2 and open-source models, improving output quality while reducing inference costs. Organisations can apply the system using a few hundred task-specific examples, converting a base model into an AI agent that generates answers, collects feedback and iteratively refines results through recursive self-improvement.
The platform also curbs unnecessary computation by determining when enough information has been gathered, ending research queries early to lower expenses without sacrificing quality.
Poetiq claims the approach delivered measurable gains, setting a record on the ARC-AGI-2 benchmark of 1,000 visual reasoning puzzles, where GPT-5.2 topped the previous best score by 16%.
“For ARC-AGI 1 and 2, we used recursive self-improvement to produce specialized agents in a matter of hours,” said Shumeet Baluja, co-founder of Poetiq.
Founded by former Google DeepMind researchers including Baluja and Ian Fischer, the company is entering a fast-growing segment for LLM enhancement platforms, as rivals such as AI21 reportedly explore multibillion-dollar deals.
