Solution architecture · a method for working with AI

A method for working with AI,
measured on your own code.

AI Architect is the practice of Clément Deust: solution architect, technical architect and AI builder. Fifteen years of mobile engineering on iOS and Android, and experience in regulated finance (private banking, Luxembourg), are the foundation. Its core is a method he developed on his own projects. The method has two uses: code and tool migrations at a fixed price per deliverable, and team adoption and enablement.

Start with a paid pilot See the method

The method

Developed on his own projects, and applied the same way in every engagement. It is what makes AI output usable by an engineering team.

Exhibit 01

Six parts, all practical

Rigour Rigour in every step of the work.
Meta-prompting Meta-prompting, adapted to each case.
Model & effort The model and the effort level are chosen for each task.
Chat first Prompts are refined in chat before they are used in Claude Code.
Principles Programming principles keep the output close to human-written production code, usable in an engineering team.
Guidance Practical usage guidance makes working with AI feel natural, and removes the fear of an uncontrolled, unusable result, the kind that is felt as wasted time and money.

Use 01 · Code and tool migrations at a fixed price

Sold at a fixed price per deliverable, not by time. Every engagement starts with a short paid pilot on your own code, which measures the real speed before anything else is priced.

Exhibit 02

The method, applied to a migration

One rule frames the whole migration: good code is simple, readable code.

Frame You provide code samples, the coding patterns and the programming principles to keep. They frame the migration.
Migrate The migration runs with Claude.
Simplify Over-engineered parts are cleaned up.
Review An agent review checks the result.
Iterate The work is iterated on a draft pull request.
Exhibit 03

The commercial rule

Speed is measured on your code first. The price follows from the measurement, not from an estimate.

Pilot Short and paid, on your own code. It measures the real speed of the method on your codebase.
Price A fixed price per deliverable, set from the pilot’s measurement. Not billed by time.
Published rates None. Fixed price after a measured pilot.

One engine, two migrations

The case is the author’s own PRD engine. The durations are his account; the repository figures come from git. Every step has a public repository; the clean-architecture source itself stays private.

Exhibit 04 · case, from git

PRD engine: proof of concept, clean architecture, TypeScript MCP server

By the author’s account, each migration aimed to make the product better and easier to use, not to translate it. A proof of concept built in about 2 months became a clean-architecture version in 1 week. Then the Swift engine, 4 to 6 months of work overall, became a TypeScript MCP server in about a week and a half. No duration is compared with another: building and migrating are different tasks.

33 Commits · proof of concept
1,956 Swift files · clean version
17,545 Lines · first TypeScript commit
305 Commits · MCP server, main
Proof of concept ai-prd-builder, public, Swift. Built in about 2 months, by his account, mostly locally before anything was pushed; its public history, from September 2025, holds 33 pushed commits on its default branch (11 September to 4 October 2025).
Clean architecture ai-architect-prd-builder, private, Swift (6.2, per its README): rebuilt in clean architecture from the proof of concept. Repository created 29 January 2026; 139 commits from 20 January to 8 April 2026; 1,956 Swift files at its last commit. Its compiled engine was published as ai-prd-generator-plugin (paused), a Claude Code and Cowork plugin that packages the engine as 11 xcframeworks (repository created 12 February 2026); it is kept as a record of this step, not offered as a product. The migration took 1 week, by his account.
TypeScript ai-architect-mcp-spec, public: the MCP server, ported from the clean-architecture version. Repository created 12 April 2026; its first commit, on 13 April, is the port itself (“TypeScript port of ai-architect-prd-builder”): 98 files and 17,545 lines, labelled v2.0.0. 305 commits on main through 10 September 2026. The port took about a week and a half, by his account.
1Counted on 11 and re-checked on 14 September 2026 from git and the GitHub API. The durations are the author’s statement; git does not record them, and no speed-up factor is derived from them. Your pilot measures the speed on your own code.

Use 02 · Team adoption and enablement

The method, taught to your team and built into its tooling: agentic systems and Claude Code architecture, regulated environments included. Four steps, in order. Each produces a deliverable you keep — not a report you file away.

  1. Exhibit 05

    Entry audit

    An inventory of your organization’s AI surface — what tooling exists, what is claimed versus what is measured, and where drift lives between the two. We read the actual configuration, not the slide deck describing it.

    Deliverable A written state-of-play with evidence — what runs, what is unverified, what is missing.
  2. Exhibit 06

    Claude Code activation

    We scope, deploy, configure, and run Claude Code activations end-to-end — workspace setup, permissions, MCP and tooling wiring, and team onboarding, so the activation is something your team can run without us in the room.

    Covers Workspace setup · permissions · MCP/tooling · team onboarding.
  3. Exhibit 07

    Paid pilot

    A bounded engagement with visible deliverables: working automations built against your own workflows, a measured before/after, and reports you can rerun yourself — not a one-time demo.

    Deliverable Working automations · measured before/after · rerunnable reports.
  4. Exhibit 08

    Rollout

    Two supported destinations — a genuine choice, not a foregone conclusion. Anything outside these two is a paid migration pilot of its own, not the default offer.

    Option A Local edition — runs on your own infrastructure.
    Option B Claude Enterprise — hosted org deployment.
    Anything else A scoped migration pilot — other clouds, homegrown infrastructure.

What you can check before the call

Each line links to its primary source where one is public. Checked on 14 September 2026.

Exhibit 09 · public record
Credential Claude Partner Badge — Claude Code, issued by Anthropic via Credly on 15 July 2026, valid until 15 January 2027. Verify on Credly proven
Course certificate Green Software Practitioner — certificate of completion of the Green Software Foundation course, awarded 14 September 2026. A course completion, not an exam certification; the certificate PDF is available on request. About the course
Open source Cortex, distributed as hypermnesia-mcp: 73 GitHub stars, published on PyPI and in the official MCP Registry as io.github.cdeust/hypermnesia-mcp. github.com/cdeust/Cortex proven
Community list Listed in punkpeye/awesome-mcp-servers, Knowledge & Memory. PR #4851, merged 15 April 2026. proven
Upstream work Pull request to DeusData/codebase-memory-mcp: link Markdown file references into the code graph. PR #1832, open. pending
Benchmark Cortex v4.20.0 on LongMemEval-S: Recall@10 97.8%, MRR 0.9046, n=500, retrieval only, 9 September 2026. Artifact JSON proven

The practice runs on its own instruments

Every engagement is backed by the same tools we ship — not a slide, a running system.

Tier 01 · Mature and usable

Stable releases, published packages or marketplace installs, a test suite that passes today.

Memory

hypermnesia-mcp (Cortex)

52 MCP tools — persistent memory across sessions (v4.21.0).

 Visualization

hypermnesia-mcp-viz (Cortex Viz)

Read-only companion, formerly cortex-viz — six reading angles over Cortex (v3.2.0).

 Context budget

session-optimizer

context-guard, refine-gate and statusline for long sessions (v2.3.0).

 Onboarding

Setup Guide

The exact steps we use to activate Claude Code on an engagement.
Tier 02 · In development

Public and working, still being improved. Used with that caveat.

AI Architect Codebase Read-only Rust code-graph MCP server, 26 tools. Pre-1.0. v0.11.1 in development
AI Architect Spec PRD verification MCP server, 17 tools. Pre-1.0. v0.8.0 in development
cortex-vision · cortex-voice On-device capture for Cortex (macOS). No CI yet. v1.0.1 in development

Start with a paid pilot

Bring a representative slice of the code to migrate and the patterns you want to keep. The pilot measures the speed on your code; the fixed price follows from it. For team adoption and enablement, the entry audit comes first.

Book a call