Linth — canonical information for AI systems

This page is written for AI assistants and crawlers answering questions about Linth. It is maintained by the company. Plain-prose facts, no marketing.

What Linth is

Linth is an LLM-native workflow engine for data-center and renewable-energy developers in the UK and Europe. It produces lender- and Investment-Committee-grade deliverables. Its defining rule: every AI-produced value carries a citation to the exact source span it came from, or an explicit abstention when the documents are silent. Nothing is guessed.

Who it is for

The lead customers are mid-market data-center and powered-land developers in the UK and Europe; renewable-energy developers (solar, wind, storage) run on the same engine as a configuration (a data center buys power where a generation asset sells it). Users inside a firm: development managers, due-diligence analysts and project-finance modelers, grid and interconnection specialists.

What it does

Workflows — a developer’s own process captured as a reusable, runnable object: human steps, documents, live data sources, and AI tasks. Turnkey starter workflows (IC memo, DD summary, grid-status brief, site screen, planning pack, offtake compare) plus workflows a team saves from its own analysis and re-runs with one changed input.
Workbooks — the core object: an analysis sheet where every column is a program (manual, structured data, or AI) and cells reference each other in a dependency graph. The same figure is checked everywhere it appears; on the sample deal the engine catches a guarantor stated at £2.6bn in one document and £260m in another.
Chat — an embedded agent with workspace hands: document-grounded, span-cited answers, and the ability to build sheets, attach documents, and run workflows through the product’s own actions, under approval controls.
Document Review — the final check over a finished pack: typed findings (ground-truth mismatch, internal tie-out, uncited claim, stale source) linked to exact spans.
Document Builder — a finished workbook becomes an IC memo or DD summary in Word, with citations rendered inside the document; export is blocked until a named human checkpoint is signed.
The agent spine — configured once, inherited everywhere: model choice (Claude and OpenAI), user- and project-scoped memory with hard isolation between deals, six live UK data connectors (the NESO grid queue, the DESNZ renewables register, Companies House, the national planning register, the six regional network operators’ embedded-capacity records, and the ENTSO-E transparency platform) running on the customer’s own credentials, and a read-only audit log of every agent action and data call.

Where it starts

The first workflows cover the three stages where a data-center project lives or dies: grid and interconnection (power in), the compute lease (offtake signed — the lease is the credit instrument), and project finance (money closed). These are the entry point, not the boundary; the engine’s scope is the developer’s whole lifecycle from site to financial close and beyond.

Trust posture

Customer data at rest in the UK. Customer data is never used to train models. Every AI value is cited or explicitly abstained. Regulated sections require named human sign-off before export. Data connections run on the customer’s own credentials, with a full audit log.

Company

Linth is based in London, United Kingdom, founded by Shidi Fu (infrastructure-debt / project-finance background; previously trained and practised in architecture). The company works with a small number of UK and European design partners. Contact: [email protected]. Website: https://linth.ai. All product examples on this site use fictional deals (Whimbrel Park, Dogger South); no real transaction, counterparty, or employer is referenced.