Adrian Bertagnoli is a founding engineer at Callosum developing heterogeneous intelligence: AI systems that coordinate specialized models, agents, and processors instead of assigning every task to the same model and hardware. His work addresses a central infrastructure challenge for agentic AI: matching each subtask with the combination of intelligence and computing power it actually requires.
Bertagnoli studied at ETH Zürich and co-authored 2021 research applying reinforcement learning to image-guided atrial-fibrillation treatment, with affiliations at ETH Zürich and King’s College London. He subsequently contributed to computational-neuroscience research on the origins of grid cells, working with researchers affiliated with MIT and Stanford to test whether proposed explanations of spatial representation withstand biological evidence.
He later joined Callosum’s founding engineering team, describing the move as an opportunity to optimize algorithms and computing hardware together in a public announcement. His infrastructure work centers on three related ideas:
- Heterogeneous recursive language models: Agents investigate long documents as external environments using Python, keyword searches, and regular expressions, then route subtasks to different models and processors. On the OOLONG benchmark, Bertagnoli reported a Cerebras configuration that was seven times cheaper and five times faster than its comparison baseline.
- Task-specific visual web navigation: Combining Qwen3-VL-8B-Instruct with Kimi K2.5 or GPT models separates demanding visual and textual reasoning from inexpensive browser actions such as zooming, reducing the resources required for routine operations.
- Automated model-and-hardware routing: An orchestration layer estimates task complexity and selects suitable model-and-processor combinations, advancing beyond manually assigned routing rules.
His AI Engineer Europe presentation places these approaches within a broader shift toward coordinated specialization across model architectures, agent workflows, and physical infrastructure. Callosum is also developing a co-located heterogeneous compute cluster with CommonAI, supported by a $2.9 million ARIA grant.
