India's Sports Authority wants to spot elite talent faster — and a new research framework aims to automate the part that human scouts actually do well.
Researchers have built a multi-agent system that combines MediaPipe's skeletal tracking with Meta's Llama-4-scout vision-language model to assess athletes during mass recruitment drives. The system follows Sports Authority of India protocols and evaluates qualitative markers — spinal articulation, form degradation, fatigue — that simple repetition counters miss entirely. A "Smart Grid" chunking strategy breaks video into 3x3 temporal segments, cutting computational overhead by more than 88% without losing the movement continuity that matters for coaching judgments. A self-correction loop has each output cross-checked against quantitative metrics before anything gets written to storage, which is the paper's answer to the hallucination problem that dogs every LLM in high-stakes contexts.
The RAG layer is the detail worth watching. Coaches can query the system in plain language — "find athletes with high endurance but poor core rigidity" — instead of writing SQL against a rigid schema. That shift from structured queries to semantic search is small on paper but meaningful in practice: it puts the analytical tool in the hands of coaches who think in physiology, not databases.
The system is novel, but it is also a research prototype aligned to one national body's protocols. How it generalizes beyond SAI's specific assessment rubrics — or holds up when a scout tries a query the authors didn't anticipate — is the question the paper doesn't yet answer.