Research

Research areas we're invested in as an AI infrastructure company. We're not a frontier-model lab — we don't train foundation models, and we don't claim to publish empirical AI research at that level. What we do bring is the infrastructure layer's perspective on questions the field is actively working on: what gets observed, governed, and made safe at the platform layer, between the application and the model.

These essays are our positions, with real citations to the work we draw from. Each closes with the questions an AI infrastructure company is well-placed to take on, and an invitation to collaborate.

Areas

AI SafetyWhat safety properties can the platform layer enforce that the model layer can't reliably guarantee?

Explainable AIAt the platform layer, explainability is audit + decision provenance — not "why the model said X" but "why the platform allowed X."

Context Window ManagementHybrid long-context + RAG + semantic cache. When does each pay back, and what does the platform owe each pattern?

RAG PerformanceMCP-as-orchestration, reranking, retrieval evaluation — and why an AI proxy is the right place to measure end-to-end.

Multi-Agent EvaluationsA2A protocols and the eval substrate that drops out of having audit, cost, and trace data on every step.

Self-Learning PlatformWhat if the platform's configuration learned from its own operational data — routing weights, risk-score weights, cache thresholds — under bounded drift?

---

For shipped engineering whitepapers describing how Scutum is built today, see Whitepapers. For a cross-cut of open problems we'd like to take on (from each whitepaper's "What we don't know yet" sections), the relevant sections are Risk-Bounded Remediation §"What we don't know yet" and Cost-Aware Routing §"What we don't know yet".

If you're working on any of these, write to [email protected]. We collaborate where there's mutual benefit and share data from design-partner deployments under research-collaboration agreements.