Compyle
Menu

The platform

One platform behind every agency we own

Each agency keeps its name, its clients and its people. Behind them sits one platform, built once and shared. It trains the team, speeds up delivery, takes the repeat work out of the week and runs the back office.

Four pillars on one foundation

Five modules do the day-to-day work. Three more sit underneath them and hold the whole thing together.

Peoplelearn

Training that fits each person's role, clients and sector.

Engineeringship

Pipelines, evals and AI tooling that take code from idea to release.

Operationsautomate · run

Find the manual work, automate it, and run the back office once.

Growthgrow

Better quotes, and new AI services to sell to clients.

your agencylearnpeopleshipengineeringautomaterunoperationsgrowgrowthcore · trust · insight

People · learn

Training built for each person

Most training is generic. learn starts from what the agency already knows: its niche, its clients, the rules those clients work under, and each person's role. Then it writes the training to match.

  • Newsletters, short video lessons, articles, podcasts and briefings, picked for one person
  • Critical actions when a rule or a client's needs change
  • A skills matrix for each role, with the gaps shown plainly
  • Certification tracking, for example cloud and security
  • Compliance training with signed attestations, kept as evidence for regulated clients
  • Onboarding paths for new starters
  • A client briefing before every project starts
Illustration of an open book with three cards rising from its pages: a video lesson, a pair of headphones and a certificate.

learn

The learn module: a personalised feed for a delivery lead, with a critical action, a video lesson, a client briefing, a skills view, certifications and an onboarding path.
A feed for one delivery lead, built from her clients, her sector and her role.

Engineering · ship

From ticket to release, with AI checked at every step

ship is the delivery engine. It fits the languages and tools your team already uses. It adds AI where it saves time, and tests that AI as strictly as any other code.

a change to a prompt, a model or retrievalchangeevals✓✕✓✓gate 0.92releaseblockedcost per task1.9p→0.4ponly if quality holds

Pipelines

  • CI/CD that fits the agency's existing stack
  • AI code review and test generation, kept or rejected by a named engineer
  • Security scanning: code (SAST), dependencies, secrets and a software bill of materials (SBOM)
  • A preview environment for every branch
  • Infrastructure as code, so every environment can be rebuilt the same way

Evals and cost

  • Eval suites for each use case, built from golden datasets of the agency's real tickets and outputs. In plain terms: a fixed set of real examples with known good answers.
  • Model-graded rubrics, calibrated against human labels before they are trusted. The AI marking the work is itself checked against people first.
  • Agent evals for tool-call accuracy and task trajectories. They check that an agent took the right steps, not only that the answer looks right.
  • Evals as CI gates: every prompt, model or retrieval change runs its evals and is blocked if quality drops. Nothing gets worse without someone noticing.
  • Shadow and canary releases for prompts and models, not only code. A change is tried on a small share of work before everyone gets it.
  • Cost-aware routing: a cascade tries the smallest model that passes the evals and escalates only when confidence is low. Cheap work stays cheap.
  • Prompt caching, semantic caching and batch inference for work that isn't urgent. Repeat questions and overnight jobs cost less.
  • Distillation: repeat work moves to small open-weight models, fine-tuned with LoRA on the agency's own examples. A small model trained on your work can match a large one at a fraction of the cost.
  • OpenTelemetry GenAI tracing for the cost, speed and quality of every task. Each client and project shows its real AI cost.

Hybrid edge AI

  • Inference hardware installed in the agency's own office
  • Quantised open-weight models served through vLLM or llama.cpp. Smaller, compressed models that run well on one machine.
  • Cloud sync, or cloud overflow when the queue is long
  • Client data kept on the premises for regulated clients

Agent tooling

  • MCP connectors to the agency's tools. A standard way for agents to reach Jira, GitHub and the rest.
  • Guardrails and personal-data redaction
  • A person approves anything that reaches a client
the agency's officeedge nodeclientdatacloudoverflow only,when alloweddata stays on the premises

ship: evals and cost

The ship module's eval gate: a pull request blocked because one rubric failed, with quality scores by rubric, cost per task, the routing cascade and a trace of one task.
A cheaper model is blocked because one quality score fell below the gate.

ship: edge AI

The ship module's edge view: an on-premises node with its health, GPU memory, requests and cost per thousand tokens, the models it serves and the data rules for each client.
Models running in the agency's own office, with data rules set per client.

ship: delivery

The ship module's delivery view: a new-project screen offering starting points such as booking and appointments, online shop and client portal, next to a table of projects and their status.
New projects start from a tested starting point, such as a booking system or a client portal.

ship: agents

The ship module's agent queue: work drafted by software agents with keep, edit and reject actions, and a record of who approved what.
Agents draft reviews, tests and support replies. A named person keeps, edits or rejects each one.

Operations · automate

See how work really flows, then take out the manual steps

Every agency has work done by hand that nobody has had time to fix. automate finds it, puts an hours figure on it and offers a tested automation for it.

  • A map of the agency's processes, from enquiry to invoice
  • Process mining from system events: tickets, commits, time entries and the CRM
  • Team-level views only. No keystroke logging and no screen watching, in line with the ICO's guidance on monitoring workers
  • Automation gaps, each with an estimate of the hours it would save
  • A library of ready-made automations, each with a cost estimate based on usage
  • Bring your own key (BYOK), so AI usage is paid at cost price with no mark-up
enquiryquoteset-up−18 hbuildreport−26 hinvoice−14 hfrom tickets, commits, time entries and the CRMteam-level only: no keystrokes, no screens

automate

The automate module: a process map from enquiry to invoice with manual steps picked out, a table of automation gaps with hours saved each month, and a proposed automation with its cost.
The process map shows where work is done by hand, and how many hours an automation would save.

Operations · run

The back office, run once for every agency

Finance, payroll and HR take up a lot of a small agency's week. run moves them to one shared team, so managers get that time back.

  • Finance and invoicing
  • Time recording
  • Payroll and HR
  • Contracts, procurement and insurance
  • Resourcing and capacity across the group, with a shared bench so engineers can be lent between agencies
Illustration of a tidy back office: filing drawers, a calculator, a stack of documents and two wall clocks, with one drawer picked out in green.

run

The run module: finance, payroll, people, contracts and insurance tasks handled by a shared team, with what the agency needs to do.
Finance, payroll, HR, contracts and insurance handled once, centrally.

Growth · grow

Better quotes, and new services to sell

grow is the commercial side. It helps an agency price work well, and gives it production-ready AI services to offer its own clients.

  • A CRM and a shared price book
  • Scoping and estimates drawn from past projects
  • Proposal drafts, edited and signed off by a person
  • AI accelerators for clients: RAG assistants, chatbots, agentic workflows and document automation, built to regulated-industry standards
  • A client portal for status, approvals and reports

The accelerators come from the founder's work at AWS, building these systems for enterprise customers in regulated industries.

Illustration of a glasshouse with rows of potted plants at different stages of growth, from seedlings to a tall plant.

grow

The grow module: a quote being built from call notes, with each line matched to a price book item and a sign-off step.
Notes from a first call become scope lines matched to a shared price book. The managing director signs off every quote.

How it connects

Three modules sit under the rest. You rarely see them on their own, but they are what make the others work.

Foundation · core

Connectors and the agency's knowledge

Connectors to the tools the agency already uses, such as GitHub, Jira, HubSpot, Xero, Slack, Microsoft 365 and Google Workspace. They feed a knowledge graph of clients, projects, people, code and contracts. That graph lets learn personalise, automate find gaps and grow estimate. It also powers "ask the agency": search and answers over the agency's own knowledge.

clientsprojectscodecontractspeoplecoregithub · jira · hubspot · xero · slack · microsoft 365 · google workspace

Foundation · trust

AI governance and compliance

A register of every model and agent, each with a named human owner, audit trails and a usage policy. Readiness for ISO 27001, SOC 2, Cyber Essentials Plus, UK GDPR and the EU AI Act, with DPIAs where they are needed. Data kept separate for each agency and each client. Access reviews and vulnerability management. This is what lets an agency win regulated clients.

agentownerchecktriage✓review✓tests✓replies✓audit trail

trust

The trust module: an agent register listing each agent's model, human owner, access and approval rule, with compliance readiness and staff attestations.
Every agent, its model, its owner and what it can reach.

Foundation · insight

Margin and performance in one view

Utilisation, delivery health, hours saved, AI spend by client and by project, and early signs that a client relationship needs attention. The screens on this site show hours rather than money.

one viewutilisationhours saveddelivery healthAI spend by client

insight

The insight module: estimated against actual hours by job and by kind of work, with the lead's explanation of each difference and suggested price book changes.
Estimated hours against actual hours, by kind of work, so the next quote is closer to the truth.

AI at cost price

Agencies bring their own model keys. AI usage is billed by the provider at cost, with no mark-up from us.

ship's routing and caching keep that bill small. insight shows exactly where it goes, by client and by project.

A week with the platform

One ordinary week at an agency, and where each module appears in it.

This week

The this-week view for Quillmere Digital: projects in delivery, agent work waiting for a person, quotes and back-office items, and a day-by-day list of tasks, each with a named owner.
One week at an agency. Every entry names the person who owns it.

See it for your agency

We'll walk you through the platform and talk about how it would fit your team and your clients. A first conversation is private and commits you to nothing.

Start a private conversation