Models, knowledge and data,
all inside your network.

Agents, knowledge bases and models run on servers the university or company owns. Off-site access goes through one gateway, and whether to use any external model is your decision.

Where everything runs

Every component sits inside your network boundary. External models are optional and off by default.

Off-site accessStaff and students via browser or app
GatewayVPN · single sign-on · access log
Your network · campus or company
Agent runtimePlans tasks, calls tools, asks for approval, keeps records
Local modelsOpen models served on-site, swappable
Knowledge baseSlides, papers, policies, project files
Business systemsAcademic, internship, OA, as needed
Logs & auditEvery call, approval and edit can be replayed
External model APIOptional. Only called if you allow it; can be redacted or switched off

Six deployment principles

We follow these on every project. Each design follows your environment, data boundaries and acceptance criteria.

Identity & roles

Single sign-on, with role-based limits on what each person and each agent can see.

Least privilege

Agents reach only the systems and data listed in their job description.

Human approval

Grades, admission results and writes to business systems always need a person.

Untrusted by default

Instructions inside web pages, files or messages are never executed directly, as a guard against prompt injection.

Data stays home

Materials and logs are stored locally and never used to train outside models.

Full audit trail

Every call, approval and edit is recorded, replayable and auditable.

Every agent
has a job description.

Before going live we write down who it is, what it knows, how it works, what “done” means and which systems it may touch.

Digital TACourse · students
Role
A course TA that practises with students
Knows
This course's slides, question bank and lab manuals
Rules
Hints first, explains later, never hands over answers
Delivers
Every session logged to the learning record
Access
May run Verilog simulation; cannot change grades
Matching assistantSchool · academic office
Role
Drafts recommendations and reallocations
Knows
Project profiles, student records, school rules
Rules
Hard constraints first; every pick needs a reason
Delivers
A recommendation list with conflicts and risks
Access
Read-only records; both sides must confirm
Literature assistantLab · PaperTree
Role
The lab's paper librarian
Knows
The lab's paper library and reading notes
Rules
Every claim traces back to a source
Delivers
Review drafts and citation maps
Access
Visible to lab members only
Knowledge assistantCompany · internal
Role
Answers staff questions on policies and projects
Knows
Internal docs, policy library, project files
Rules
Answers only with sources; says when unsure
Delivers
Sourced answers and follow-ups
Access
Content filtered by department

How we work

No need to replace everything at once. Prove one workflow first, then expand.

  1. Define

    Choose one workflow; define its boundaries, success criteria and where people must approve.

  2. Connect

    Connect materials, systems and permissions; write the agent's job description.

  3. Pilot

    Run on real tasks, with your team reviewing every step.

  4. Operate

    Monitor, iterate and update; apps and sites keep shipping, platforms keep running.

Questions

What servers do we need?

It depends on model size and concurrent users. We recommend a configuration for your scenario, and pilots can often start on servers you already have.

Which models can we use?

By default, open models that run on-site. If a task needs a stronger commercial model, it can be connected with your consent, with data redacted, and switched off at any time.

Will our data be used for training?

No. Materials, conversations and logs stay on your servers and are never used to train outside models.

How do you stop agents being misled by malicious instructions?

External content is untrusted by default: instructions in web pages, files or messages are not executed directly; agents can only reach the systems in their job description, and key writes need human approval. Before launch we test specifically for prompt injection and privilege escalation.

Start with one real workflow.

Tell us what you want to solve and we'll work out where to begin.