Internal proof project · Built by Nocta/case/gorolli

GoRolli.

A logistics workflow tool we built end-to-end to validate our automation and agent patterns under real operational load. We use it as a proving ground — not as a product we sell.

Internal proof project  Live · internalNot for sale
Context — § 01

Why we built it.

Logistics is a domain where the gap between “demo-quality AI” and “runs every hour without breaking” is brutal. We wanted a workload of our own to harden the patterns we use in client engagements: queueing, retries, observability, agent tool-use, and human-in-the-loop overrides.

GoRolli was the answer to that. Built, deployed, and run by Nocta — for Nocta.

Role
Designed, built & operated by Nocta
Year
Status
  Live · internal
Audience
Internal operators only
Type
Proof project · not a service Nocta sells
Problem — § 02

What was hard.

Manual dispatch

Orders, statuses and exceptions were moving by hand across a chain of tools. Each step lost context.

Flaky integrations

Vendor APIs were inconsistent. Errors needed to be caught, retried and explained — not just logged and forgotten.

No source of truth

Operators couldn't answer “where is this?” without paging through three systems and a chat.

What we built — § 03

A small operating system for logistics work.

Workflow engine

Typed workflow steps with retries, idempotency keys, and human-overridable transitions. Every event is auditable.

Dispatch agent

A constrained LLM agent that drafts route plans and notifications, with tool-use limited to vetted internal APIs.

Operator console

A single view of in-flight work — what's blocked, what needs a human, what the system did unattended.

Evaluation harness

Golden cases, replays, and drift checks. We can re-run yesterday's traffic against today's prompts and code.

Stack — § 04

Boring infrastructure. Current models.

Frontend
Next.js · Tailwind · React Server Components
Backend
Node services · queue workers · typed RPC
Data
Postgres · object storage · event log
LLM layer
Provider-agnostic adapters · tool-use API · cost meter
Evaluation
Golden suites · replay runner · per-step latency & cost budgets
Ops
Structured logs · per-trace IDs · paging on golden-suite regressions
What this proves — § 05

The patterns we use in client work — pressure-tested by us, first.

Operational

Agent + workflow systems can run quietly day-to-day with a single operator — if the seams are built right.

Engineering

Evaluation harnesses pay for themselves the first time a model upgrade silently changes behaviour.

Product

Operators trust systems they can override. We design overrides in from day one.

Next — § 06

If you have a workflow that looks like this — talk to us.

GoRolli is not a product we sell, but the patterns inside it are exactly what we build for clients. Start with an audit.