I direct AI to build systems from scratch: databases, APIs, queues, deployment, monitoring, knowledge layers, agents. Came up through customer support at Payoneer, Fiverr, and Bancor, so I notice when something was built without thinking about the human using it. Everything here was built this way. My tools are Claude Code and Codex.
I run 2,300+ Claude Code sessions a month, orchestrating parallel builds across multiple projects. I think in systems: what needs to exist, how the pieces connect, what the user actually needs. Then I direct AI to build it, verify it works, and ship it.
AI automation platform for Israeli small businesses. 15+ microservices, queue based processing, row level security, tenant isolation. Hebrew first, aimed at operators who do not use apps. Working on Noa, a humanized Hebrew agent with vertical specific prompt graphs and low confidence escalation. Wired Langfuse, OpenTelemetry, and Prometheus for per turn agent tracing. MCP servers on both sides: one exposes Ottomatt tools to external agents, another runs mechanical quality gates that block delivery on recurring failure modes.
Designed and built for a customer support scenario where ticket volume outruns the docs team. The system ingests documentation nightly, retrieves with semantic search, escalates through a 68 branch decision graph when confidence drops, and feeds corrections back into the next embedding cycle. The point was an agent that gets sharper every night, not staler. 150+ endpoints, 26 step nightly pipeline, three self learning loops. Full architecture in the PDF below.
Israeli real estate intelligence. Neighborhood scoring engine aggregating crime data, school quality, transit access, building permits, and price history. Multi signal weighted scoring across public data sources.
DeFi analytics and agent tooling for stake.link. JaJ, a RAG chatbot with prompt injection filters and L1 to L4 confidence auto-calibration, lives at stakedotlink.money/LLM. MCP server exposing 15 on chain tools to external AI agents (Chainlink oracles, Morpho vaults, LinkPool subgraph) is live at stakedotlink.money/agents. Plus Telegram community monitor, Morpho vault dashboards, automated reporting.
Built the AI automation platform from scratch. 15+ microservices, tenant isolated Postgres with row level security, Hebrew first WhatsApp surface, automated research pipelines, agent self improvement feedback. Working on Noa, a humanized Hebrew agent with vertical specific prompt graphs. Wired full stack observability through Langfuse, OpenTelemetry, and Prometheus. Built MCP servers on both sides, including mechanical quality gates that block delivery on recurring failure modes. Directed the full stack using Claude Code and Codex as the primary development motion.
Leadership course for young managers and entrepreneurs in Israel. Practical management frameworks, a peer cohort of operators, and access to senior executives, paired with depth and values from Jewish wisdom on leadership. I am there to sharpen the business and product side of how I build AI tools.
Built the SDL MCP server exposing 15 on chain tools to external AI agents. Shipped JustAnotherJay, a DeFi chatbot with RAG retrieval, prompt injection filters, and confidence tier auto calibration. Analytics dashboards and Telegram community monitor.
Multi channel customer experience (Zendesk, Discord, Telegram) for 50K+ users on a complex DeFi platform, through 10x user growth. Designed escalation workflows, routing rules, CSAT loops, and the knowledge base content model that fed product decisions. This is where I learned what knowledge systems actually need from the consumption side.
High volume B2B marketplace support. Dispute resolution, fraud analysis, complex case management at scale. Contributed to the knowledge base and response playbooks used across the team.
Enterprise cross border payments. Disputes, fraud investigation, gateway case management for merchant partners. Cross functional process improvement and knowledge base authoring.