Build with llamay

By sector

Eighteen sectors, the constraint that actually bites in each, and the pipeline people build once they have accepted it. The architecture underneath is the same everywhere: the model runs where the data already is.

The same shape, eighteen times

Finance

What bites: material non-public information, and a supervisory expectation that you can say where a document went. Built: filings and research notes summarised against a house template; trade-narrative drafting from structured fills; KYC document extraction under a schema so the fields land in the same shape every time. Why local: a draft of an unannounced deal is exactly the text that cannot be posted to a vendor endpoint while it is still unannounced.

Healthcare

What bites: identifiable patient data, and a minimum-necessary rule that makes "we sent it somewhere and they deleted it" the wrong shape of answer. Built: clinic-letter drafting from structured notes; coding assistance that suggests rather than assigns; de-identification as a first pass before a research extract leaves the estate. Why local: the de-identifier is the one component that must see identifiers, so it is the one component that cannot be somebody else's.

Pharmaceuticals and life sciences

What bites: trial data under blind, and a documentation trail that an inspector will read years later. Built: protocol deviation triage; adverse-event narrative drafting from case forms; literature screening where the screening criteria are themselves confidential. Why local: an unblinded read that leaves the estate is not a privacy incident, it is a trial integrity incident.

Defence and national security

What bites: classification, and networks that are deliberately not connected to anything. Built: report summarisation and entity extraction on disconnected estates; multilingual triage where the languages are not the commercially popular ones. Why local: see air-gapped — the binary is static, the transfer is a file and a checksum, and -pull=false removes the write routes so the deployment cannot be asked to fetch anything.

Legal

What bites: privilege. A privileged document that transits a third party is an argument you do not want to have. Built: discovery triage and responsiveness ranking; clause extraction against a playbook; first-pass summarisation of a matter's correspondence. Why local: privilege is about who could have seen it, not about who did.

Insurance

What bites: claims contain medical and financial detail at once, and decisions have to be explainable to a regulator and to the claimant. Built: first-notice-of-loss structuring; policy-wording comparison across versions; fraud-signal shortlisting that a human then works. Why local: the volume is large and the per-record sensitivity is high, which is the worst combination for per-token pricing.

Government and public administration

What bites: residency rules, procurement, and records law that assumes you can produce what was processed. Built: correspondence triage and routing; plain-language rewriting of official notices; translation for service delivery in the languages actually spoken locally. Why local: "the data stayed in the building" is a sentence a records officer can sign.

Education

What bites: student data protection, and budgets that do not survive a per-seat subscription. Built: formative feedback on drafts; question generation from course material; accessibility rewriting at reading level. Why local: a school can run a 3B on hardware it already owns, and the marginal cost of the next thousand essays is electricity.

Manufacturing

What bites: the useful text lives on an OT network that is separated from the corporate one on purpose. Built: maintenance-log triage; work-instruction drafting from engineering change notices; failure-mode clustering across shift reports. Why local: the machine that holds the logs is frequently the machine that is not allowed to reach the internet.

Energy and utilities

What bites: critical infrastructure rules, and field sites with intermittent connectivity. Built: inspection-report structuring; outage-narrative drafting from SCADA events; regulatory filing assistance. Why local: a substation laptop that works when the link is down is worth more than a better model that does not.

Telecommunications

What bites: subscriber data is regulated separately from everything else, and volumes are enormous. Built: ticket deflection and routing; network-alarm summarisation; churn-reason extraction from free-text notes. Why local: at this volume the unit economics decide it before the privacy rules do.

Retail and e-commerce

What bites: nothing much, which is why this is a good place to notice that the argument is not only about regulation. Built: catalogue enrichment; review summarisation; support deflection; search query rewriting. Why local: cost per token at catalogue scale, and latency in front of a search box where a hundred milliseconds is the whole budget.

Logistics

What bites: documents arrive as scans from hundreds of counterparties in a dozen formats. Built: bill-of-lading and customs-document extraction; exception triage; carrier-message normalisation. Why local: the parsing is high volume, low value per document, and runs fine on a small model — which is the profile per-token pricing suits worst.

Media and publishing

What bites: embargo, and sources. Built: transcript structuring; archive search over a back catalogue; headline and summary drafting under house style. Why local: an embargoed draft and a source's identity are both things that should not be in anybody else's logs.

Agriculture

What bites: connectivity, and seasonal peaks that make a subscription awkward. Built: agronomy-note structuring; compliance-record drafting; advisory translation for a workforce that does not share one language. Why local: the field has no bars, and the season has no patience.

Real estate and construction

What bites: long documents, many parties, and a contract chain where the exact wording is the product. Built: lease-clause extraction; RFI and submittal triage; site-report structuring from photographs' captions and notes. Why local: commercially sensitive terms, and files too large to enjoy paying per token for.

HR and recruiting

What bites: candidate data is personal data, and automated decisions about people attract specific rules in several jurisdictions. Built: CV structuring into a schema; interview-note summarisation; job-description rewriting for clarity. Why local: keep the data inside, keep a human deciding, and keep the model's role to structuring rather than ranking.

Scientific research

What bites: unpublished results, and reproducibility. Built: literature screening; methods-section extraction; lab-notebook structuring. Why local: a pinned model file and a fixed seed are reproducible in a way an endpoint that silently updates is not — and unpublished results are unpublished.

Humanitarian and non-profit

What bites: beneficiary data can be dangerous in the wrong hands, and the operating environment often has no reliable network. Built: needs-assessment structuring; multilingual intake; report drafting for donors from field notes. Why local: in some contexts "this never left the laptop" is a protection measure, not a compliance checkbox.

What is common to all of them

Three things recur, and they are worth naming because they are the actual reasons rather than the stated ones.

Where to go next

Recipes has the pipelines named above, with code. Deployment patterns covers air-gapped transfer, multi-tenant isolation and the local-first fallback. By industry goes deeper on the five most regulated of these.