Logistics: Ontime runs its development on Claude, with the methodology inside its own codebase

Context

Ontime is a national logistics operator that, alongside its own operation, runs BPO services for third parties. Like any company carrying twenty years of systems, it faces two problems at once: a legacy Oracle APEX platform to be modernised module by module, and operational processes that still live in a shared mailbox and in the heads of the people who handle it.

The operational starting point was concrete: a Microsoft 365 mailbox as the official channel for pickup requests, shipment tracking and delivery incidents, handled entirely by hand by a team of eleven agents. No formal allocation of the load, no view of the real state of any email, and zero data traceability: no shipment could be reconstructed back to the email, attachment, row and approver that originated it. The company’s own Quality and Audit function had begun to require exactly that.

The constraints ruled out the easy answers. The transport management system cannot be changed and has no API documentation. Email stays the channel: customers are not pushed onto a new portal. Automatic replies are forbidden on complaints. Personal data cannot reach an external model without redaction, and the deployment has to operate under EU AI Act obligations with mandatory human oversight. A generic assistant was never an option.

Collaboration

Thinkia works with Ontime on two fronts, and on both the engine is Claude, Anthropic’s AI.

The methodology as a Claude Code plugin

Thinkia’s delivery method — AI-SDLC — is not explained in a document: it is executed. It reaches Ontime packaged as a Claude Code plugin, with the workflows available as commands inside the development environment itself. Each workflow loads its step sequence, validates its inputs against the project tree, executes each step through a version-pinned skill with a declared output contract, and closes in a gate. A blocking finding halts the run and persists nothing. Agents produce; humans sign.

On the Oracle APEX platform

That plugin carries 3 commands, 20 skills, hooks and two MCP servers connecting Claude to the customer’s own Oracle and Jenkins. Quality does not rest on good intentions: 64 eval cases across 15 suites run gated in continuous integration, and a prompt-hash ratchet forces re-evaluation before merge as soon as a prompt changes. One skill even carries a hard model gate: it aborts if the active model is not Claude Opus, and it has an eval case verifying that it does not give way even when the user is in a hurry.

In operations

The same method produced the full specification and design of a governed agentic system for the mailbox: a deterministic core owning validation, registration and audit, with the AI and the orchestration as replaceable layers around it. Every external system — mail, transport management, street directory, audit log — is reached through an isolated least-privilege MCP server. A guardrails pipeline sits in front of every model call: PII masking, prompt-injection detection, caching and per-call audit. The principle is written into the architecture itself: the AI proposes, the rules decide. No model output reaches the transport system without deterministic re-validation, and every action lands in an append-only SHA-256 audit log.

And it did not stop at delivery

The toolkit is installed in Ontime’s own development environments and their teams are trained on it. Today it is Ontime’s developers who run Claude-assisted development as standing practice: same commands, same skills, same evaluation gate in CI. Thinkia extends the skill set per module instead of doing the work.

Outcome

In production at Ontime, 12 Ontime developers have the toolkit installed and in use — Claude-assisted development is the team’s standing practice on their Oracle APEX platform, not a pilot and not a trial. More than 400 APEX modules and screens have already been through it, at 80% less time per module than under the previous method. Consumption runs on Ontime’s own Claude accounts, not Thinkia’s: it is theirs, day to day, and it continues between engagements. Because the method travels with the codebase, each new module enters the same gate without a Thinkia session, so the capability stays with the customer.

Measured in the operations delivery — a Thinkia measurement, verifiable in the project artefacts — the same method produced 10 capabilities and 17 features decomposed from Ontime’s requirements, with a full specification set for the lead feature and a data design with migrations verified against a real engine. 28 workflows were executed and closed under gate, each with a recorded verdict, and zero gates were signed by an agent: every open consideration is carried forward as explicit debt rather than quietly cleared. Three navigable role-based prototypes, produced from the specifications themselves, let Ontime’s supervisors and executives walk the flow before anything was built — all across two working days of Claude Code sessions, with Claude Opus as the active model.

What is not yet a result, stated plainly: the mailbox system is specified and designed, not deployed. The automation, accuracy and handling-time figures are Ontime’s own targets against a baseline still to be established. This page does not present them as achievements.

Case metrics correspond to specific projects and do not constitute a promise of results; each context produces its own range.