AI that does
real work.

Replen is a supply-chain planning skill that sits between a general AI and a real operating decision. It can size safety stock and reorder points, diagnose forecast error, handle intermittent demand, compare EOQ against MOQ and pack constraints, and segment a portfolio with ABC–XYZ. The more important part is the gate before the number: which service-level definition do you mean, is the variability measure valid, and what data is still missing?

I built it by turning repeated planning decisions into a written answer protocol, moving the arithmetic into five dependency-free Python calculators, adding reusable CSV inputs, and separating public planning knowledge from private work context with a red-line check. It is a live pilot for decision support — no ERP write-back and no invented data.

This synthetic distribution company tests one transaction across the full operating chain: customer demand, available stock, purchase orders, split confirmations, inbound receipt, quality release, customer delivery, supplier invoice matching and cash. Seven exception stories deliberately try to break the flow.

The result is measured in assertions, not a demo script: 27 checks across the scenario suite and 12 more in the second end-to-end realism pass. Every ERP action is simulated; no production tenant or real organisation data is connected.

Six specialist planners cover demand, S&OP, supply, inventory, production and logistics. A supervisor holds the end-to-end objective. A separate verifier tries to refute important outputs instead of quietly fixing the work it grades.

The key mechanism is a task-by-task trust ladder. An agent may be trusted for a read-only inventory check while remaining draft-only for a purchase recommendation. Payments, commercial commitments and consequential external communication stay with a person.

Two private operational tools sit behind this public case study. An offline MRP calculator turns planning spreadsheets into grouped replenishment recommendations. A voyage and invoice review tool connects purchase orders, packing lists, invoices and receipt history to an auditable payment decision.

Both tools work from ordinary PDFs and spreadsheets, keep processing on the device and stop before the ERP. The operational implementations remain private; this page contains only generic workflow descriptions and no real transactions or organisations.

An org chart for a single life: a chief-of-staff agent on top and specialists underneath — investing research, personal brand, health, and the day-job planning stack. Every agent drafts; one human reviews and decides. It is how this site, the research notes and a good share of my working week actually get produced.

A visual report arguing that AI does not arrive in a company as one event — it climbs a staircase. S0: humans relaying everything by hand. S1: agents swapped into existing seats, same org shape — most companies stop here. S2: the plateau — shared world-state and gated writes, the realistic ceiling for an existing company. S3: the AI-native company, born that way rather than converted. Each step ends with two take-aways — one from the employee’s seat, one from the owner’s — and none of the three walls between the steps are technical.

Two animated information-flow maps of one distribution company. Mode A swaps each department for an agent: the old shape with new internals — orders, stock and invoices still relayed hand-to-hand, humans as approval gates. Mode B is AI-native: one core optimiser with small agents at the edges, all three flows solved at once, humans only at the two ends. A seven-dimension table pins down where the evolution actually happens.

A match-logging and hitting-partner app for club tennis players: find someone at your level, set up a hit, submit the score, watch the season build. Designed, built and shipped to production by one person — a supply chain planner, not a software engineer — with AI doing the heavy lifting: React 19, Supabase, Vercel.