ArgonIQ
A white-label aftersales platform for machinery manufacturers, starting with Italy's. The manufacturer lists the machines installed at each customer's site and gives every customer a command center for their equipment: the latest technical documentation in as-built condition, grounded and cited answers about setup, parameters, maintenance, spares and issues, and a warm handoff to the service team when a person is needed.
One front door for every installed equipment issue: guided resolution when it is safe, a pre-qualified service case when it is not.
- Market
- Italian machinery manufacturers
- Role
- Founder & Technical Lead
The problem
Italy is Europe's second machinery producer: roughly €50B a year across some five thousand mostly family-owned, export-heavy firms. For these manufacturers, aftersales is already about a quarter of revenue, and with margins on new machines getting sharper, service is where competitiveness and customer loyalty are defended.
Yet digital services barely exist in the sector. The day-to-day reality of aftersales is phone calls, email chains and the senior technician's personal WhatsApp: no audit trail, no knowledge control, and expertise that walks out the door with every retirement. Most technician vacancies in Italy are already hard to fill. The answers usually exist, inside 400 page PDFs nobody opens while a line is down.
A command center for their customers
ArgonIQ runs under the manufacturer's brand. Their customers get one place for everything about the machines they own.
The installed base, listed
The manufacturer registers each customer’s machines by serial. Everything on the platform resolves to that machine’s as-built configuration, not to a generic model.
Documentation, digital and current
Manuals, service procedures, drawings, test reports and certificates, always the latest approved revision, available for the life of the machine.
Questions answered from the documents
Machine setup, parameters, fine tuning, spares, maintenance, issues, quality. Answers are grounded in the documentation and cite the exact source page. When nothing supports an answer, it says so and hands over.
A ticket that arrives half done
When the issue is not solved on the spot, the customer opens a ticket prefilled with the serial, configuration, symptoms, checks already performed and likely causes. Service desk integration available. The discovery and introduction calls mostly disappear.
Recurring problems, surfaced
Existing tickets are mined per machine family. Recurring issues become structured feedback for engineering and R&D instead of folklore in the service inbox.
Not every document is for every reader
A machine builder's documentation ranges from public manuals to service bulletins to margin sheets. Every document lands in one of four visibility tiers, and the tier decides what the AI may do with it.
- T1
Customer visible. Standard documentation for the machine family: manuals, procedures, certificates. The assistant may cite it in answers.
- T2
Customer specific. As-built documents for that customer’s serials: drawings, configurations, test reports. Citable, visible only to the customer they belong to.
- T3
Internal service knowledge. Service bulletins, known failure modes, repair notes. They shape the diagnosis and the ranking of likely causes, but never appear in a customer reply.
- T4
Restricted. Costs, margins, design internals. Ingested through a separate path that the customer-facing system cannot query at all.
The hard problems
An AI that advises people around industrial machinery has to earn trust structurally. Most of the engineering went into making the safe path the only path.
- 01
Safety is decided in code, not by the model
Stored energy, live electrics and interlocks force escalation in plain code before any AI reasons. The model can raise the severity of a situation, never lower it. AI drafts service procedures; people publish them, with a two person sign-off on safety steps.
- 02
Cited or silent
Every customer facing claim resolves to a page in the manufacturer’s approved documentation. If an answer cannot be grounded, the system escalates instead of guessing.
- 03
The tier wall is structural
Tier visibility is enforced at the database with Postgres row level security and checked again on the generated text, not requested politely in a prompt. A T3 phrase reaching a customer reply is treated as a release blocking defect.
- 04
A model per task
Claude Haiku classifies, Sonnet writes the customer answer, Opus reasons through the hard diagnostics. Cost follows the difficulty of the task.
Why now
The new EU Machinery Regulation (2023/1230, applying from January 2027) makes digital instructions first class: documentation has to stay available, printable and current for the machine's lifetime, which for industrial equipment means decades. Keeping every document digital, versioned and tied to the exact serial is what the platform does by construction, so it directly supports that compliance work.
And the market is moving: most machinery manufacturers expect digital services to be decisive for machine sales within a few years, while today those services are nearly nonexistent in the sector. For an Italian OEM competing on service quality rather than price, this is the gap.
Where it stands
Early build, ahead of its first pilot. The platform runs end to end on seeded data: cited answers, the prefilled ticket flow, the tier isolation probe, and an audit trail that lets every AI decision be reconstructed with its sources and safety classification. Every release replays a set of golden fault scenarios through the real pipeline in CI; one unsafe answer, internal leak or cross tenant read blocks the release. Current violation count: zero.
The design target is to contain 20 to 50 percent of level 1 support contacts, measured against the manufacturer's own baseline during a 90 day pilot on one machine family, not promised up front.
Stack
- Product
- Next.js 15, React 19, TypeScript, Tailwind CSS 4, tRPC v11, Turborepo
- AI
- Anthropic Claude (Haiku, Sonnet, Opus), Voyage embeddings, hybrid retrieval with rank fusion, eval harness in CI
- Data
- PostgreSQL with pgvector, Drizzle ORM, row level security for tenants and knowledge tiers
- Platform
- Better Auth, CASL default deny authorization, S3 / R2, OpenTelemetry, Playwright and Vitest