A logistics brokerage ran dispatch, clients and vendors on spreadsheets and manual handoffs. A custom ERP normally means a 6-month, multi-developer build.

Orchestrated 4 AI coding agents plus a PM agent across isolated git worktrees to ship a full ERP — role-based access, row-level data isolation, a Kanban operations board.

Production ERP delivered in weeks, not months — 80% test coverage and 15 end-to-end flows.

Dispatchers hand-planned delivery routes every morning while field managers logged workdays on paper — hours lost, wasted fuel, no real visibility.

A two-part platform: a VRP routing engine (OR-Tools) that turns an address export into optimal daily routes, plus a mobile PWA where field managers track meetings, mileage and KPIs in real time on a custom design system.

Daily routes built in seconds instead of hours; field operations visible on one admin dashboard with Excel/PDF export.

Developers and staff kept re-asking the same questions and digging through scattered internal docs — every answer cost a senior engineer’s time.

A retrieval-augmented (RAG) chat assistant over the internal knowledge base, wired to NotebookLM through a custom MCP server — deployed and guarded by an automated security-regression suite.

Instant, around-the-clock answers from trusted internal sources, freeing senior engineers from repeat questions.

Monthly reports were assembled by hand in Excel — hours of error-prone copy-paste every month by non-technical staff.

A native desktop app (Tauri + Rust) where a non-technical user drags in the raw .xlsx and a Rust engine processes and writes the finished report — no install ceremony, no spreadsheets touched.

A recurring multi-hour manual task reduced to a single drag-and-drop that anyone on the team can run.

Staff needed route distances and map links but worked entirely inside their ERP — switching to external tools and copy-pasting addresses was slow and error-prone.

A stateless FastAPI microservice the ERP calls directly: it parses addresses, computes real route mileage via OSRM and returns a map link — plus a geocoding pipeline (self-hosted Nominatim) that resolves address exports to coordinates.

Routing and geocoding available on demand inside the system staff already use — no new app to learn.

Several bots ran on a server with no monitoring — a silent crash could go unnoticed for hours, and “AI-built means fragile” was the standing objection.

A Telegram monitoring bot that polls log files in real time, batches alerts, detects error spikes and crashes, and posts a daily digest — with mute and quiet-hours controls.

Failures caught the moment they happen instead of hours later — proof that AI-built systems can be production-solid.