Skip to content
CPLT
Request a call
Menu

AI assistance & automation / Customer-owned builds

I build AI assistants and automate repeated work.

I build AI systems around your documents and processes.

What you get

  • Answers from your own documents, with the source page shown for each answer.
  • Paid work has a written scope and a price agreed before it starts.
  • You keep and operate the finished system.

What you don't get

  • I do not promise savings or an accuracy rate.
  • I do not sell model subscriptions or seats.
  • There is no retainer requirement.
Request a free call See the evidence ↗

Free 45-minute remote call.
A one-page note on fit and next steps.

What happens after the free call

  1. 01

    Written note

    A one-page written note on fit and next steps.

  2. 02

    Paid trial, if needed

    It is a separate paid engagement with a written scope and a price agreed first.

  3. 03

    Comparison

    A comparison of current work and trial results at the same quality standard.

  4. 04

    Decision

    A decision to continue, revise or stop.

AI-generated portrait of Stéphane Lepain in a navy suit

Stéphane Lepain
Founder and builder · CPLT
How I work →

Fictional source check

Interactive example, not a live AI assistant. Fictional text, not a customer result or a CPLT service promise.

A useful answer shows its source.

Routine request guide · p. 2

Keep completed request records for 30 days.

A named reviewer checks each draft before it is sent.

Illustrative document / 02

30 days. Source: p. 2.

Two more examples

A named reviewer. Source: p. 2.

Not supported by this example. The source does not say.

Choose a question to highlight its source.

01Evidence, with limits

Read what worked. And what did not.

Public test / 13 October 2026

AI answers, checked against GitLab's own.

I hid GitLab's published CAIQ answers and tested an AI workflow using only its public handbook.

174/236held-out questions matched
GitLab's published answer

174 matched · 2 differed · 60 abstained

The always-Yes baseline matched 225 of 236 labels. The workflow correctly answered 0 of 11 No/N/A cases. Both wrong answers passed the quote check unflagged. A genuine quote does not prove the answer is right. The expert signs off every answer. This is a public-document test, not a finding about GitLab's security or a promised accuracy rate.

How the routing figures were counted

176 of 236 answered. 63 of 236 sent for expert review. The 47 questions used for tuning are excluded. The review rule changed after the held-out run; these routing figures re-score the same answers.

Read the test, sources and failures ↗

My working system / September–October 2026

One person prepares what specialists sign off.

The figures, their limits, and a failure I caught

I prepare alone the review that a team of 3 to 5 specialists signs off: security architects for each domain, a programme manager and a product owner.

2 years
in my day-to-day work, one person
2,773 pages
read across 38 versions, Sep 2026
50/50
verdicts matched in a 50-case test
15 checks
run on every review before I see it

Duration: my own statement. Page count, verdict match and check count: measured, September to October 2026. My own working system, not a customer case study.

I built this for my own professional work. It is not a CPLT deployment. Read a failure and how I caught it →

Full evidence and limits ↗

Discuss one recurring task in a free call ↗

02The work

What is one repeated task costing you?

How to count it

Finding an answer, preparing a routine update, copying information between steps, checking a draft before it goes out.

Count how many people, how many hours a week, at what they cost per hour. That is the number a new way of working has to beat.

01

An internal knowledge assistant ↗

I build a workspace around your documents and processes. It can answer questions with source pages and prepare drafts for a person to check.

02

Automation for repeated steps ↗

I scope how information moves from a request to a checked result. Connections, permissions, exceptions and approval before actions are agreed for your task.

03

Optional private hosting ↗

When your requirements justify it, I build a workspace in your environment. I assess data access, operating responsibility and cost before choosing where it runs.

Discuss one recurring task in a free call ↗

03The engagement

Agree what to build and how to check it.

How an engagement works, and my working notes
  1. 01

    Describe today

    Record the effort, delays, errors and cost of completed work.

  2. 02

    Agree the checks

    Name the reviewer, acceptable results and cases that need a person.

  3. 03

    Bound the trial

    Agree examples, scope, written price and criteria before paid work starts.

  4. 04

    Decide what follows

    Compare the results and full cost. Continue, revise or stop.

Is it worth the cost?

I compare a sample of real tasks, done today and done with the tool, at the same quality. Failed attempts count too.

Cost per accepted result ↗

Decide what the tool may do.

Before building, I agree what information may be used, which services may receive it, who approves actions and who handles missing or conflicting information.

Fictional source check and limits ↗

I scope it, build it and walk you through it.

I deliver paid projects with a written scope, price and checks. You work directly with me. I use an open-source LibreChat workspace, a LiteLLM gateway and Ollama for local open models.

More about how I work ↗

Discuss one recurring task in a free call ↗

04Reading & resources

Notes on AI cost, choices and checks.

More notes
Compare self-hosted AI with API costs ↗GDPR Article 28 and data-processing checklist ↗What Is an Automation Engineer? ↗All essays ↗RSS feed ↗
Watch the 2 minute 46 second video, or get the task questionnaire

Watch / 2 minutes 46 seconds

Put AI to work. Make the business case.

I explain how to choose one task, agree human checks and compare the full cost.

Made with AI: the presenter uses my AI likeness and voice. The examples are fictional, not customer results. Selectable captions are available in the YouTube player.

Watch with the full transcript ↗

The player connects to YouTube when loaded. Playback starts when you press play. Watch on YouTube ↗

Prepare for a conversation

Start with one task, not a tool.

Two fillable pages, free to download. Save it and email it to me, or bring it to the call. Leave out confidential information.

Download the free task questionnaire ↗

Discuss one recurring task in a free call ↗

05Before you start

Questions about working with CPLT

What can CPLT build for my business team?

I build custom AI assistants for questions, summaries and drafts using permitted information, and automations for repeated steps. I agree the task, connections, acceptance checks and human review before paid work starts.

What happens during the free scoping call?

The free 45-minute remote call starts with one recurring task, how often it happens and what you want to improve. You receive a one-page written note on fit and next steps; any paid work has a separate written scope and price agreed first.

Do I need private hosting?

No: private hosting is optional, and approved external APIs are another delivery choice. I assess the data boundary, actual processing routes, operating responsibilities and full cost before choosing the setup.

How do you assess whether an AI trial is worthwhile?

I compare representative work against agreed acceptance checks and count implementation, model/API usage, hosting, human review, correction, maintenance and support, including failed attempts. Time released is capacity; I do not promise cash savings, payback or an accuracy rate.

What happens after handover?

I document operating procedures and walk your team through the runbook, with responsibilities agreed before building. Ongoing support has its own agreed scope and price.

Start here

Tell me what keeps coming back.

A free 45-minute remote scoping call, ending in a one-page written note on fit and next steps.

EU-based · Remote delivery · Written scope first

Request a free call