For a decade I studied Talmud twelve hours a day before I wrote a single line of code. It turned out to be the best debugging training I could have asked for.
Today I'm a Full-Stack Engineer building production systems in an elite technology unit — React and TypeScript on the front, Node.js, Express and MongoDB behind it, shipped end to end.
What sets me apart isn't the stack. It's that I take the problems that get handed back. More than once I've cracked a challenge experienced teams had already given up on — by learning the thing properly instead of working around it, and by connecting small, seemingly unrelated details into a pattern nobody else saw.
Next: AI/LLM engineering — RAG, agents and model-based systems in production, on a strong full-stack foundation.
🔒 Private repository — client work, source not public.
End-to-end development: React 19 + Vite + Tailwind (Frontend), Node/TypeScript + Express + MongoDB (Backend). Secure video LMS (signed-token playback, per-user moving watermark, progress tracking), online donations via payment gateways, courses & certificates, full i18n (Hebrew/English, RTL), and WCAG AA accessibility compliance (Standard 5568).
Result: A cost-optimized, multi-provider media architecture that keeps the nonprofit at near-zero cost.
🔒 Private repository — source not public.
Built from scratch a self-hostable central authentication server in TypeScript, Express, MongoDB: Google OAuth2 with PKCE, JWT over a JWKS endpoint with encrypted auto-rotating keys, refresh-token rotation with reuse detection, RBAC, session management, NIST 800-63B password policy, and a full audit log — at 100% test coverage (800+ tests).
Result: Deployed to production, serving real applications as a central auth service.
Everything below is public — clone it, read it, run it.
| Repo | What it is |
|---|---|
| codegen | A fully-local autonomous coding agent on Ollama. One engine behind a CLI, a browser GUI, a VS Code extension and a Telegram bot — CPU-friendly, no API keys |
| docrag | Layered local RAG for Hebrew/English documents: classify → extract → SQL or hybrid vector search, behind a router. Offline, stdlib + numpy + SQLite |
| rag-financebench | A RAG pipeline measured against the FinanceBench dataset across several evaluation dimensions |
| skills | Claude Code Skills — modular, auto-loaded instruction units for LangGraph agents, MCP integration and knowledge-graph memory |
| Repo | What it is |
|---|---|
| template-service | Production-ready Node.js service template — Express, strict TypeScript, MongoDB, Zod, Winston, Helmet, Docker, Vitest |
| Website-Monitor | Uptime monitor that polls a list of sites and records only status transitions, with a React dashboard |
| minio | S3-compatible file storage — React frontend, Express API, MinIO and MongoDB, wired up with Docker Compose |
| bun-template | Microservice template on the Bun runtime: Hono, tRPC, MongoDB, Zod, Pino |
- 🔍 Retrieval — taking
docragpast brute-force cosine: ANN indexing and a reranker pass on top of the hybrid dense⊕BM25 layer - 🤖 Agents — extending
codegentoward multi-step planning and safer tool execution, still fully local - 📊 Evaluation — measuring RAG quality against standard benchmarks
(
rag-financebench) instead of eyeballing answers - ☁️ Production — moving model-based systems off a local Ollama box and into deployed services
