Industrial data that
earns its keep.

A general-purpose LLM answers on its own word — and grades its own work. In our trials, that failure shows up most where it hurts industrial data most: numeric values. Weights, power ratings, dimensions come back plausible, confidently stated, and wrong — and a plausible wrong number looks identical to a right one. That's why InRecipe separates generation from verification: every value a skill proposes is checked against a cited source by an independent validator before it reaches your database. Every field comes back a verified value, an honest gap, or a flagged estimate routed to a human. The system would rather refuse than write something that might be wrong.

Every populated field carries its evidence: which source it came from and when it was observed. Values that can't be verified don't ship as facts. In the marine pilot's ground-truth checks, every datasheet extraction we hand-verified came back right — and where the enrichment disagreed with the customer's own master, it was the master that turned out wrong on 3 of 4 disputed values, with the manufacturer's evidence attached.

You get an honest gap — a real blank, distinct from "didn't look." Where a bounded estimate is possible, it's capped low, flagged as an estimate, and routed to a human review queue. Never a silent guess.

Your records, outputs, and review history stay in your own isolated engagement tenant. LLM skills run via a commercial API where inputs and outputs are not used for training. And when the opt-in web tier searches for a datasheet, the search provider sees only a public manufacturer-and-model query — never your records.

The rule follows ownership, not policy. If a manufacturer or regulator already publishes it — a model number's specifications, a customs classification — it can become shared, curator-verified knowledge for your vertical, and you inherit what earlier customers' engagements already built, from day one. Everything that's about your operations — your records, enriched outputs, service history, your in-house names and conventions — stays in your tenant and is never shared. And nothing moves to the shared layer automatically: every promotion is curator-reviewed and verified against the manufacturer's own sources first.

The mess is the starting point, not an obstacle. Production runs have handled mixed Finnish-English registers where the same machine appears as "ATLAS 90kW VSD #2" in one row and "kompressori 90 kW" in another. If your data were clean, you wouldn't need this.

You can — a data steward clears 2,000–5,000 records a year, different stewards disagree, and the work restarts as new records arrive. And hand-edits leave no trail: when someone types a value into the ERP, there's no record of where it came from, why it was chosen, or what it replaced — six months later, nobody can say. In production runs, InRecipe enriched close to 2,000 items in hours, every change traceable, dated, and attributable, with documented evidence on every field — and it doesn't forget what it learned. And speed is a dial, not a limit: enrichment fans out across parallel agents — up to 200 of them when a deadline demands — while every record still passes the identical validator.

The machinery could. An enrichment layer with evidence and an independent validator isn't industrial-specific, and product catalogs everywhere suffer the same blanks, duplicates, and unverifiable values. But the value doesn't come from the machinery alone — it comes from vertical depth: curated manufacturer catalogs, model-number grammars, domain vocabularies, documented failure modes. That's why we're deliberately focused on industrial data for now — equipment, parts, item masters, supplier catalogs — where that depth already compounds. If your data is industrial-adjacent, ask. If it's a retail catalog, we'll tell you honestly: not yet.

One business case that matters, a 50–100 record sample (a CSV export is fine), a couple of manufacturers or suppliers, and an hour or two with someone who knows the data. No integration, no software installation. We bear the costs of the first week.

Pilots run on exports: CSV in, structured evidence-graded output back, in a format your systems import. Deeper embedding — new records enriched automatically as they're created — comes at the Autopilot stage. And for agentic workflows, an MCP server is on the roadmap: InRecipe as a verified-data tool your AI assistants and copilots can call directly, with the evidence traveling along.

Not a per-seat license. Pricing follows what you choose to enrich: how many records, which fields and how demanding they are, and the level of verified quality you need. The free pilot week calibrates all three on your own data — so the quote is grounded in your register, not a rate card, and you know what fixing it is worth before you commit to anything.