Founder of Real AI. Three decades across data and AI delivery at industrial scale, with a current focus on foundation models for the real world: Hominis — a family of situated, auditable, compute-aware foundation models trained on allocation time at Leonardo, the EuroHPC supercomputer at CINECA, Bologna.
Tarry Singh founded Real AI to build the European stack for foundation models that hold up under real-world deployment — situated, auditable, and compute-aware. Before Real AI, three decades of data-and-AI delivery at industrial scale across banking, energy, manufacturing and pharma. The pattern across every engagement: the systems that succeed are the ones that show their work and survive contact with regulated production environments.
The Hominis programme is the engineering vehicle for that pattern. The SYMPHONY consortium is its first proof point at the intersection of foundation models, neuroscience and robotics.
Hominis is a family of foundation models calibrated to three properties — the three pillars of the cathedral above:
Trained on the context the model will be deployed into — industrial-automation logs, EU regulatory text, multi-lingual scientific corpora — not just the open web.
Every output traceable to a substrate region; every adaptation to a task token. Compositional control surfaces, bounded behaviour, external evaluation built in.
Trained on EuroHPC allocation at Leonardo / CINECA. Per-paper tCO₂e reporting; absolute compute-budget caps declared in the DMP.
As coordinator, Real AI runs the consortium spine: project management, financial reporting, dissemination, regulatory watch, the Consortium Agreement, the Data Management Plan, and the open-science discipline. The work is invisible when it goes well, which is the point.
On the science side, Real AI integrates the contributions of the three other partners into a single substrate: ingests Newcastle’s neuromodulatory framework as the O2 mechanism, threads CREATE’s haptic shared-control formalism through as O3’s task baton, and trains on UP Robotics’s industrial-automation demonstrator codebase for O1 and O4.
Build an automated pipeline that ingests a software system and emits a four-layer representation (structural, behavioural, historical, rationale) over a single graph. Threshold: coverage ≥ 90 % of functions and ≥ 80 % of inter-module dependencies on the two demonstrator codebases.
Demonstrate, on a pre-registered evaluation protocol, that the SYMPHONY substrate outperforms three named baselines: (a) a frontier LLM agent, (b) a best-in-class static-analysis + knowledge-graph pipeline, and (c) an LLM + RAG baseline — on a held-out benchmark of 200 engineering-task instances. Threshold: ≥ 20 % relative F1 improvement on task-relevant-subgraph recovery and ≥ 15 % in expert-rated actionability.
Beyond the science, Real AI and UP Robotics have agreed in principle on a joint venture as the primary exploitation vehicle for the SYMPHONY substrate. Terms are set out in the Consortium Agreement at grant preparation; the JV term sheet is targeted for execution by M40. Newcastle and CREATE retain academic IP and grant the JV a non-exclusive, royalty-bearing licence for industrial use. Read the full go-to-market and IP plan in the proposal.