Learning-based control
Routing systems
We study signal learning, model selection, and inference policy so routing becomes a learnable, optimizable, and auditable ML problem rather than a pile of hand-written rules.
Open GitHubResearch
Ongoing frontier research for enterprise AI.
We treat routing, reasoning, caching, protocols, and multi-agent runtime as foundational systems work for enterprise AI.
RoutingReasoningRuntimeProtocols
Frontier technical research turned into deployable infrastructure for enterprise AI.
Research Focus
We use frontier machine learning and systems research to make AI deployable in the real world.
We care not just about stronger models, but about learned decision mechanisms and runtime foundations that let AI operate safely, reliably, and sustainably in production.
Capability, cost, and boundaries
Reasoning and efficiency
We study adaptive reasoning, semantic caching, and efficiency frontiers to learn better trade-offs among capability, latency, and cost, making stronger models truly deployable.
Deployment, protocols, and runtime
Runtime and protocols
We study runtime isolation, multi-agent execution, and open protocol layers that give models, tools, and enterprise systems a stable way to work together in production.
Publications
Papers, protocols, and open work.
Across routing, reasoning, caching, runtime, and protocol design, we publish papers, drafts, and open implementations.
arXiv 2026Featured paper
vLLM Semantic Router: Signal Driven Decision Routing for Mixture-of-Modality Models
A system treatment of signal-driven routing for mixture-of-modality systems, covering semantic policy, model selection, and controllable inference paths.
Open publicationWhen to Reason: Semantic Router for vLLM
Routes harder questions to stronger reasoning models instead of paying the same reasoning cost on every request.
Open publicationCategory-Aware Semantic Caching
Uses dynamic cache thresholds and TTL policy to improve reuse while staying aligned with query type.
Open publicationSemantic Inference Routing Protocol (SIRP)
A protocol proposal for classification-aware routing across AI infrastructure layers.
Open publicationResearch Method
Research, code, and systems practice move together.
We keep papers, open implementation, and protocol design in one loop rather than advancing them in isolation.
Frontier problems
We focus on routing, reasoning, runtime, and control as technical problems with real deployment consequences.
Open implementation
Research is grounded in working systems, from signal extraction and decision logic to DSL compilation and provider-neutral runtime behavior.
Protocol lens
We turn system insights into reusable interfaces and protocol proposals that can push broader infrastructure forward.
Continue
See products and architecture.
See how the research becomes a deployable system.