Summary
who I am in one paragraph
Software Architect (Principal Engineer at SAP) shipping software since 1987 — from OS/2 system programming to autonomous AI-agent fleets. Today: production ML services for expense-fraud detection at SAP Concur, plus applied LLM/agentic engineering — I design and operate a fleet of autonomous AI coding agents shipping 20–30 hands-off releases per day across 9 real products, and I teach AI-assisted development from prompt engineering to full agent fleets. Three SAP patents along the way.
AI Expertise
from teaching the models to teaching the humans
Agentic engineering
Built Maestro — Go orchestrator driving parallel coding agents through issue → git worktree → PR → review gate → auto-merge; 20–30 hands-off releases/day, 9 products. Personal 6-agent assistant fleet with tiered semantic memory (RAG, embeddings).
Classical ML in production
Fraud & anomaly detection, document and line-item classification (ONNX), OCR post-processing — SAP Concur audit pipeline at multi-region scale (US / EU / APJ / US Public Sector).
AI-assisted development
Advanced daily "vibecoding" practice: Claude Code, Cursor, Codex, OpenCode, Antigravity CLI — system-prompt & CLAUDE.md authoring, custom skills, MCP servers, subagent workflows, context management, agent-failure debugging. Author of an internal pattern library.
Prompt & context engineering
Production prompt design for code agents and LLM pipelines — extraction, classification, summarization — with structured output, tool-use design, and an evaluation-and-iteration discipline.
LLM platform economics
Multi-provider routing (Anthropic, OpenAI, Google, Fireworks, local Ollama) with cost-aware fallback chains, OAuth/API brokering, spend RCA and hard caps; self-hosted LiteLLM/OpenWebUI gateway.
Teaching & enablement
Author of AI-101 — multilingual EN/RU/HE course taking non-programmers from zero to automating real workflows with Claude Code + MCP (live case: accountant automating payroll & PayEm→SAP). 10-module curriculum up to agent fleets.
Experience
almost 40 years, reverse-chronological
SAP Labs Israel
Software Architect / Principal Engineer, Development Expert — SAP Concur
- Architect and operator of production ML audit services processing expense receipts at scale for global Concur customers: ml-audit, fake-receipt detection, tax-compliance extraction, policy audit, line-item classification — Python, AWS EKS/SQS/DynamoDB/S3, ONNX, FluxCD
- Incident ownership and RCA for multi-region production (US / EU / APJ / US Public Sector Cloud); PagerDuty on-call for ML-pipeline fanout alerts
Software Architect — Innovation Center Network Israel
- Owned Web & App architecture end-to-end: took the product's UI from zero to production and led its full-stack team hands-on
Senior Developer — SAP Portal
- Full-stack development on SAP Portal; co-invented the real-time messaging framework (patented)
Teleweb Services
Senior Software Engineer
USBPhone
Senior Software Engineer
CreoSoft.net
Project Manager · Sr. Software Engineer · Team Leader
Inopsys
Senior Software Engineer, Team Leader
Admiral Makarov National University of Shipbuilding
Systems Programmer
- OS/2 system programming and administration; later built CGI pipelines generating HTML for the university's internal hyperweb, plus AutoCAD / AutoLISP engineering automation
Independent Engineering
selected public work, 2024 – present · github.com/BeFeast · self-hosted on own Proxmox/TrueNAS infrastructure
AI-agent orchestrator: picks up GitHub issues, spawns isolated coding agents (Claude/Codex/Gemini) in git worktrees, drives PR lifecycle through review gate to auto-merge. Runs a 9-project fleet hands-off — ~30 releases/36h sustained.
Audio stem separation (Demucs) + tempo detection → DAW-ready .dawproject. FastAPI / Celery / PostgreSQL + Next.js.
Any song URL → isolated vocals + instrumental + synced lyrics: BS-Roformer stem separation on serverless GPU (RunPod), Whisper + LRCLIB lyrics alignment; React + Web Audio. Built almost entirely by the Maestro agent fleet — ~30 hands-off releases in 36 h.
YouTube → transcript → summary pipeline: Whisper transcription + LLM summarization chain, on-demand GPU (vast.ai), Chrome extension; React + Python.
Telegram-first AI agent as a single 18 MB binary — multi-provider routing, skills with safety audit.
Cross-platform media player: .NET WinUI 3 (Windows) + Rust/GTK4 (Linux), auto-update pipeline. GPLv3.
apertune (chromatic tuner), S1ADelay (modulation delay) — CLAP / VST3 / AU.
Hands-on multilingual course: non-programmers → automating real workflows with Claude Code + MCP.
Beyond Code
evidence of systems thinking
5-node Proxmox cluster + TrueNAS, 45+ containerized services, Cloudflare edge — secrets management, monitoring, independent backup domains.
Recording engineering, DSP, DAW ecosystems — domain depth behind RehearseKit and the plugin work.