Research

SmartInfer builds vertical AI applications backed by research in memory, retrieval, verification, efficient inference, and decision intelligence. The selected work below frames the technical ideas behind our product architecture.

• Verifying Structured Retrieval with Laplacian Coherence
  Submitted to NeurIPS. A label-free method for estimating retrieval confidence using graph-local coherence over evidence candidates, designed for memory systems where retrieved facts must be verified before use.
• Untrusted Builders, Trusted Gates: A Verification Architecture for Agentic Software
  Submitted to Onward! Essays. Argues that agentic software needs independent verification gates: untrusted generators may propose patches and repair loops, but acceptance should be controlled by machine-checkable gates ranging from property-based tests to model checking and theorem proving.
• Position: Agent Verification Requires Typed Workflows and Bounded Commitments
  Submitted to NeurIPS. Argues that agent verification should focus on committed traces over typed workflow actions, with explicit commitment gates separating open-ended model proposals from state-changing commitments.
• The Information Geometry and Sample Complexity of Grouped Preference Optimization
  Submitted to TMLR. A theoretical analysis of listwise preference supervision, showing when larger comparison groups provide limited gains and when evaluation noise becomes the true sample-complexity lever.
• Exploration Is Bounded by Estimation: Fisher-Directed Budget Allocation under Hill Saturation
  Working paper / submission in progress. A marketing science method for allocating exploration budget under saturating channel-response curves, using Fisher information to decide where additional spend improves estimation and where it is unlikely to help.

• AI Memory Architecture for Large Language Models: From Context Windows to Persistent Intelligence
  A 3D-8Q taxonomy of how LLM systems store and retrieve knowledge, comparing neural memory modules, retrieval-augmented approaches, KV cache optimization, and disaggregated architectures.
• Agents and Agentic AI in the Era of Large Language Models
  A first-principles overview of agentic AI: what agents are, how LLMs facilitate planning and tool use, what constitutes an effective agent platform, and a comparison of frameworks and managed services.
• Inference-Time Compute Scaling in Large Language Models: A Theoretical Survey
  Formal foundations grounded in circuit complexity: when test-time compute improves capability, the generation–verification gap that enables scaling, and why verifier quality is the ceiling.

• The Evolutionary Journey of Search Engines: Architecture and Evaluations
  A concise history of modern search engines, their architecture, and the methods used to evaluate them.
• Search 2025: From Information Access to Intelligent Execution
  How LLMs are turning search from a user-facing interface into invisible infrastructure for real-time reasoning, retrieval, and execution within intelligent systems.
• The Evolution Beyond Search — From Information Retrieval to Autonomous Execution
  Part 2 of the LLMs and Future of Search series. Examines the progression of search beyond retrieval into autonomous execution.

• The Rise of Reasoning Models: A Revival of Planning in AI
  From prediction to deliberation — how modern AI is rediscovering the value of structured reasoning.
• A Concise Review of Reinforcement Learning Methods
  Foundations of RL, deep RL, and LLM-centric approaches, with practical implications for production systems.

• Enterprise AI Outlook 2026 and Beyond: Applications, Inference Optimization, and Infrastructure
  The transition from model-centric experimentation to enterprise-scale application and system deployment. Argues that beyond 2026 the binding constraints are economic, architectural, and operational rather than model capability alone.
• Architectural Advances in AI Through a Systems Lens
  From model architecture and training paradigms to compiler infrastructure, runtime systems, accelerators, and interconnect — practical AI impact is constrained by the maturity of the execution stack beneath the model.
• Notes on Model Distillation
  A review of distillation algorithms and applications — an important method for compression and efficiency relevant to small-model deployment.

• The Transformation of Ad Engines: LLMs, Foundation Models, and Agentic Systems (2026+)
  Digital advertising’s most significant transformation since programmatic buying. Details the collapse of organic CTR, AI Overviews reshaping search ads, and the consumer reckoning that constrains how much further this can go.

More technical writing: technical articles. Academic work: selected publications, DBLP bibliography, and patents.