News · Updates
Model and product improvements across the Eagle Eye suite
Motivation and problem formulation
A multi-product sports AI suite fails quietly when each SKU optimizes a local metric while the desk workflow regresses. We formulate continuous improvement as a suite-level reliability program: reduce failure modes that matter to sports professionals (wrong brand ID, brittle SoV, ungrounded agent claims, stats comparisons on mismatched bases), and keep announced capabilities aligned with what we ship and what case studies actually support.
Architecture schematic (qualitative).
System / program architecture
Improvement workstreams map to products: CV detection & valuation factorization (Sponsorlytix / Estimator); multilingual monitoring & SoV (Dantani); agent tool-use & grounding (Koba); stats faithfulness (StatGost); roadmap vision (Turtul) and desk surfaces (Bench / Terminal-adjacent). Shared principles: athletes/fans first; transparent/auditable measurement; responsible scaling; open evaluation; sport-grade reliability; global by design. Open evaluation remains an aspiration relative to fully open leaderboards — this post publishes protocols and pending result columns for the next evaluation releases.
Illustrative evaluation profile — not a published benchmark.
Illustrative evaluation profile — not a published benchmark.
Data and methods
We prioritize regression suites built from: (1) annotated frames and placements for Sponsorlytix; (2) multilingual media items for Dantani; (3) desk-task rubrics for Koba; (4) table-grounded QA for StatGost; (5) assumption-complete exports for Estimator. Case studies remain the public quantitative anchors (Chelsea 2025/26 brand values; FIFA WC 2026 scope; WTA Makarska 2022 brand values). Image-only PDFs are cited by title only when text extraction fails.
Comparison — improvement focus by product
| Product | Primary failure mode targeted | Eval style | Public status |
|---|---|---|---|
| Sponsorlytix | Missed / confused marks under blur & occlusion | Frame + factor ablation | Live + case studies |
| Estimator | Hidden assumptions in valuation exports | Audit checklist | Announced |
| Dantani | Entity bleed / SoV instability | Multilingual samples | Live dashboard |
| Koba | Ungrounded advisory | Claim–evidence audit | Announced |
| StatGost | Numeric unfaithfulness | Table-grounded QA | Announced |
Evaluation table
| Metric | Definition | Unit | Protocol | Result |
|---|---|---|---|---|
| Suite regression pass rate | Share of critical tests green pre-release | % | CI gate | Internal eval — contact sales |
| Sev-1 reopen rate | Critical incidents reopened within 30d | rate | Incident review | — |
| Cross-product consistency | Same entity facts across SKUs | % agree | Entity probe set | Internal eval — contact sales |
| Doc drift | Announced capability vs shipped behavior mismatches | count | Quarterly audit | — |
Results posture (what we cite today)
Across the suite, the large quantitative anchors we cite in news articles are case-study figures from Resources PDFs — for example Chelsea FC 2025/26 combined sponsor brand value of $751M and FIFA World Cup 2026 scoping of 104 matches / 48 teams / 16 host cities. Those anchors validate that measurement and monitoring systems operate at professional scale; they are not substitutes for detector mAP or agent win-rate leaderboards. Image-only PDFs such as Al Ain FC, BKFC Dubai, Crisis Reputation on X, Ford Middle East PR, Transfer News, and Women’s Football and the Media are cited by title only where numeric extraction is out of scope.
Improvement tickets are triaged by user-visible failure mode first (wrong brand, brittle SoV, ungrounded advice, unfaithful stats) and by severity second. Canary rollouts prefer a single product surface before suite-wide propagation when shared entity services change. We prefer fewer doc-drift incidents over faster announcement copy.
Limitations and failure modes
- Silent metric gaming if teams optimize a narrow detector score while commercial artifacts regress.
- Announcement vs ship asymmetry for Koba / StatGost / Estimator / Terminal.
- Incomplete open eval papers — principles mention open evaluation; published eval pages remain thin.
Safety and responsible use
Improvement programs should not ship “sharper” rumour generation or more aggressive scraping that violates platform terms. Reliability includes refusal quality and auditability, not only task success.
What’s next / availability
Improvements roll through existing product surfaces. For access and partnership on evaluation collaboration, use Talk to sales. Related roadmap posts cover Bench and Turtul timelines.
Appendix — glossary
Canary. Limited rollout used to detect regressions before full release.
Doc drift. Divergence between announced capability and shipped behavior.
Suite regression. Cross-product tests that catch broken shared entities or contracts.