Ariel Rubinstein

First fruit.

Featured builds

Systems I built that run in production, from finance back-offices with years of uptime to a client e-commerce platform launching soon, plus the knowledge infrastructure and personal automation that keep my own operation honest.

• Tabletops by Nomi

  Full-stack e-commerce platform · client build · launching soon

  Designed and built from scratch: a luxury e-commerce site for a bespoke table-linen studio. Next.js + Supabase + Stripe, with a shape-aware custom-dimensions pricing configurator, checkout across 25 countries, and an oversized-order quote flow. Behind it, 19 production n8n workflows run the business end to end: order confirmation, EasyPost shipping labels, Xero invoicing and payment reconciliation, and a phone-photo → product-page + Instagram-post pipeline.

  • Design to deployment, one person + AI agents
  • 19 production workflows run the back office

• BSP reconciliation platform

  Production finance system · air-cargo GSA

  Automated the monthly IATA BSP records cycle (download, extraction, reporting), replacing a ~4–5-hour manual monthly process. Now consolidating the surrounding legacy stack (15–20 macro-driven Excel files, 9 years of records across 3 companies, 3 banks, and 4 currencies) into a single Next.js + PostgreSQL platform, built with AI agents under my direction (in progress).

  • ~4–5 hours of manual work per month removed
  • 9 years of records, 4 currencies, one platform

• Airline invoice reconciler

  Production finance system · air-cargo GSA

  Python ZIP-to-PDF extraction feeding a Power Query / Power Pivot model that compares cargo billing exports against airline invoices and surfaces the discrepancies a human needs to decide on. Heroku-hosted; outputs analyst-ready decision-point reports instead of raw spreadsheets.

  • ~4 years in continuous production
  • Eyeballing spreadsheets → decision-point reports

• Job-hunt digest pipeline

  Personal automation · nightly production

  Nightly scans across job boards and ATS APIs (Greenhouse, Ashby, Lever) covering 50–200 companies per run, LLM fit-scoring against a locked personal rubric, and a ranked morning digest. Three candidate profiles run in production with a FastAPI health dashboard.

  • Runs every night, unattended
  • Deterministic pre-filters + LLM scoring rubric

• Document-intelligence stack

  Knowledge infrastructure · self-hosted

  Kreuzberg text extraction feeding a LightRAG knowledge graph with semantic search over 50+ documents, on a nightly ingest. An MCP server exposes four knowledge-base tools to Claude Code agents, so every automation runs grounded in accumulated context instead of from scratch.

  • Knowledge graph + hybrid semantic search
  • Agents query it as a first-class tool

• Telegram agent fleet

  Personal automation · self-hosted

  A fleet of Telegram bots backed by scheduled agent jobs on a home server: system monitoring, task and calendar management, daily accountability check-ins, and a content-approval gate where a human approves before anything ships.

  • Six bots, one hybrid architecture
  • Human-in-the-loop by design