TRIJET

Projects we have delivered

Twelve client projects so far, seven taken into production. The nine below are the ones Trijet can describe: client names stay under NDA, and pilots are marked as pilots.

Client projects

Figures cover each project from start to finish, including the work with the client before anything was built. Catalogue build times count from an agreed spec.

  • Knowledge search agent

    Went into production

    3months

    from start to production

    Search over the internal knowledge base of a company with 1,000–3,000 people. It answers the question instead of returning a pile of documents.

    • 85% precise answers on the client’s test questions at acceptance
    • Research that took tens of minutes, sometimes hours, now takes a few minutes
    • Three engineers, 862 commits, running in the client’s own environment

  • Quality subsystem, UK commodity trading

    Went into production

    5weeks

    one engineer, inside the client’s codebase

    Inside a UK software platform for commodity trading, we made the accuracy of document extraction something the client can check.

    • Accuracy per field and per document, with production corrections fed back in
    • Built to the client’s conventions. They kept the code and the checks

  • Payment and delivery reconciliation

    Paid pilot

    8days

    from scratch to deployment

    Incoming payments matched to deliveries, so people only look at the exceptions.

    • Two people, with tests and full project documentation
    • Every match records why it was made, so a person can audit it

  • Corporate AI workspace

    Went into production

    18months

    of continuous development

    An in-house AI workspace where employees work with the company’s own documents and systems.

    • Two branded production deployments from one codebase
    • SSO with group permissions, an audit trail and tracing of every model call

  • Agent lifecycle platform

    Went into production

    13weeks

    to production, three engineers

    Business teams set up their own agents. The platform takes care of documents, permissions and rollback.

    • Creating, changing or switching off an agent is a managed step that can be rolled back
    • Documents attached to an agent are indexed automatically

  • Ask the database in plain language

    Went into production

    3projects

    built on the same pattern

    Employees ask a question in the chat they already use and get the number back, with no ticket to analytics.

    • Works against the real corporate schema, with validation and safety limits on every query
    • Two production environments

  • Inbound mail processing

    Went into production

    4months

    of active work, 209 commits

    Operational email turned into structured records, so nobody has to retype it.

    • Four stages behind a queue: pull, clean, classify, parse
    • A failure in one kind of message never holds up the rest

  • Agent evaluation platform

    Went into production

    Agents built on different frameworks are checked the same way, and a drop in results raises an alert like any production incident.

    • Question sets, repeatable runs and one shared set of metrics
    • Scores explained so that people outside engineering can read them

Collaborations

AI products we have built and run in collaboration. The oldest put an LLM into production in 2023.

Start with one thing

Tell us which process takes up most of your team’s time. We’ll say whether we’ve built it before.

Book a callWrite to us