I take ideas from concept to validated prototype — fast — across information retrieval, RAG, and agentic AI, then help teams turn the winners into shipped systems and large-scale research.
I work independently on how systems retrieve the right context, reason over it defensibly, and act on it in real time — the connective tissue between a good model and a system you can trust in production.
Over three decades I've carried ideas from prototype through to production systems and peer-reviewed publication — with contributions to openGauss, OpenStack, and RAMCloud, and papers at venues including VLDB and ECIR. Today I split my time between independent research and consulting for teams and startups on applied AI, system design, and technology strategy.
Each started as a research question and became a working, source-traceable system. Read the write-up, or dive into the code.
A two-system architecture that hides external-data latency in real-time voice agents: a fast agent on the critical path, a slow supervisor planning in the background, and a speculatively populated pool cache with an LLM-free reranker. Prefetching hid 95.5% of latency; simple historical counts beat MCTS at predicting the next action.
Extends graph-RAG with a fourth dimension on every edge: a RelationContext capturing why a decision was made, who approved it, through what channel, and when it expires — turning a system of record into a system of decision. Now with a soft-predicate business-rules gate and an LLM-driven web ingester.
An OSINT pipeline that turns a news corpus into a scored, fully source-traceable knowledge graph. It extracts actors and events, merges the evidence, and surfaces bridges, brokers, and corroborated leads — compressing 449 articles into a 57-node map in ~50 minutes, every claim traceable to its URL.
An open evaluation framework for goal-oriented conversation agents. Instead of grading against fixed answers, it runs A/B "matches" grounded in Cialdini's principles, against a diversity-stratified benchmark and a calibrated customer simulator — comparative testing for dialogue where exam-style scoring fails.
Peer-reviewed and independent research in information retrieval, neuro-symbolic methods, real-time agents, and in-memory databases.
Speculative Context: A Pool-Based Cache for SOP-Constrained Voice Agents — sole author. A two-system architecture that hides external-data latency in real-time voice agents using a speculatively populated pool cache and an LLM-free reranker.
Three decades building distributed systems, databases, and applied-AI prototypes — and leading the teams that ship them.
Software architecture, system design, and technology strategy for startups; independent applied-AI research across retrieval, RAG, and agentic systems.
Applied research in IR, semantic search, and large-scale visual deep-understanding; architecting distributed database and networking components adopted into Huawei platforms; prototyping RAG, agentic, and visual-language systems and leading them into product.
Led the team designing and building event-driven monitoring and management systems for a large electric-vehicle charging network.
Led R&D teams and end-to-end delivery of enterprise software projects across multiple clients.
Unix infrastructure software at Intel's Israel Development Center; system integration and connectivity in medical imaging.
Available for consulting and research collaboration on applied AI, retrieval and RAG, agentic systems, and distributed-systems architecture.