Portfolio · David Potgieter

Systems built, programmes designed — and what they changed.

Agentic AI Engineering · AI Transformation · Training & Coaching

Most of this work was built first for David's own businesses, Talent Bridge and iMPLEMENTAi.ae, and runs there every day. Each case study sets out the problem, what was built, the result and the stack behind it.

David Potgieter
11 Case Studiesengineering to coaching
124+ Skillspersonally authored
124+
AI Skills personally authored
13
AI roles running two businesses
80%
Faster agent routing workflow
2 FTE
Recruiter capacity released by AI
17
Modules in the public AI course
54+
Executives coached one-to-one
Portfolio at a Glance

Eleven projects, one practitioner.

From multi-agent architecture to a public AI course and an executive coaching method — choose a project to jump to its full case study.

Agentic Architecture

FolderDesk — Human-and-AI Operating System

Business platform, 13 AI roles, 124+ Skills and MCP execution running two businesses.

Read case study
Agent Engineering

Skills Library and Discovery

A 140+ Skill catalogue with ranked discovery, so agents load only the method they need.

Read case study
Infrastructure

Multi-Environment Agent Execution

Agents acting across Mac, Windows, cloud servers and local models; 92 → 18 seconds.

Read case study
Agent Engineering

Agent and Skill Delivery Pipeline

Persistent OpenAI agents with review, deterministic state and a 36-test regression suite.

Read case study
Local AI

Local and Hybrid AI Models

Ollama and MLX on owned hardware, benchmarked and routed alongside cloud models.

Read case study
AI Transformation

AI-Assisted Recruitment Operating System

A 21-stage lifecycle that releases two full-time recruiters' worth of capacity.

Read case study
Digital Product

iMPLEMENTAi.ae — Built With AI Agents

A production website with generated offers, previews, checkout and gated releases.

Read case study
Enterprise Architecture

AI Estimation Brain (Proposal)

Source-linked cost intelligence for a major Dubai real-estate developer.

Read case study
Training

Practical Agentic AI Course

A free 17-module course taught by David AI Coach, an AI tutor built on his method.

Read case study
Coaching

Executive Career Method and Hybrid CV

54+ executives coached through a 7-step method, now encoded into AI Skills.

Read case study
Foundations

Earlier Systems Work

Security systems, corporate facilities and hands-on technology since 2004.

Read case study
Case Studies

Agentic AI Engineering

Agentic Architecture

FolderDesk — Human-and-AI Operating System

Architect and builder · iMPLEMENTAi.ae · 2025 – Present

Context

Business knowledge kept disappearing into chat threads, and every new AI tool started from zero. Two businesses needed AI that could carry real responsibility across recruitment, sales, marketing and finance without losing context.

What Was Built

  • A layered operating system: a shared business platform (Twenty), AI roles, reusable Skills, MCP execution and durable organisational knowledge in GitHub.
  • 13 AI roles — including Recruitment Director, Assessment Director, Chief Marketing Officer and Chief Financial Officer — each with a mission, responsibilities, authority limits and Skills.
  • Role identity separated from the model, host and harness, so the same role runs on ChatGPT, Claude, Codex or a local model.
  • Mission Control for human oversight, with people and agents working on the same business records.

Outcome

Talent Bridge and iMPLEMENTAi run day to day on the system, with humans and AI agents operating the same organisation under explicit ownership.

Agent Engineering

Skills Library and Organisation-Wide Discovery

Author and architect · 2025 – Present

Context

Long prompts were being rewritten for every task and every tool. Proven methods needed to become reusable, versioned capability that any capable agent could find and load.

What Was Built

  • Skills covering engineering, recruitment, assessment, CV writing, sales, marketing, finance, research, infrastructure and automation.
  • A Skills MCP discovery layer that returns a short ranked shortlist for any task, so agents load only the method they need.
  • Progressive loading: a Skill exposes deeper references only when the task requires them.
  • Clear rules for when a Skill relies on AI judgement and when it calls deterministic scripts.

Outcome

Methods improve once and every agent benefits. The same CV-writing, screening and assessment quality reproduces across tools and people.

Infrastructure

Multi-Environment Agent Execution

Architect · 2025 – Present

Context

Agents needed to act on real machines and services — a Mac workstation, a Mac Studio, Windows, cloud servers and websites — through controlled, auditable routes rather than ad-hoc access.

What Was Built

  • MCP servers exposing bounded operational capabilities on each node, behind FastAPI, OAuth and Cloudflare.
  • Dedicated actions per machine, with same-node fallbacks so work never silently lands on the wrong computer.
  • Self-hosted GitHub Actions runners for deployments and Mac-only jobs.
  • Isolated runtime profiles for separate agent identities (public example on GitHub).

Outcome

One routing and status workflow dropped from 92 to 18 seconds and from 20 to 5 tool calls: 80% faster with 75% fewer steps.

Agent Engineering

OpenAI Agent and Skill Delivery Pipeline

Architect · iMPLEMENTAi.ae

Context

Custom AI Skills built for clients had to be produced, reviewed and published reliably, without a human supervising every step.

What Was Built

  • Persistent OpenAI Agents with hosted execution environments and reusable builder Skills.
  • Independent model review, validation and immutable publication through the OpenAI Skills API.
  • Deterministic workflow state with atomic work claiming, recovery, replay protection and controlled provider fallback.

Outcome

A repeatable build–review–publish pipeline, protected by an automated regression suite.

Local AI

Local and Hybrid AI Models

Builder · iMPLEMENTAi.ae

Context

Not every task needs a cloud model. Privacy, cost and speed called for capable models running on owned hardware.

What Was Built

  • Local model environments with Ollama and MLX / oMLX on Apple silicon, including an always-on Mac Studio.
  • Benchmarks comparing local models against each other and against cloud models for specific task types.
  • Hardware-aware routing between local and cloud models, including small local decision systems.
  • Continual-learning experiments with held-out evaluation and retained local model weights.

Outcome

Routine classification and decision tasks run privately on local hardware, keeping cloud models for work that needs them.

Case Studies

AI Transformation & Digital Products

AI Transformation

AI-Assisted Recruitment Operating System

Designer and operator · Talent Bridge

Context

A recruitment practice built on more than 10,000 personal interviews ran on experience held in people's heads, which limited how far it could scale.

What Was Built

  • A 21-stage lifecycle from vacancy intake through sourcing, screening, questionnaires, assessment, interviews, shortlisting and post-placement follow-up.
  • Reusable AI Skills for each stage, connected to the ATS, email and WhatsApp.
  • Human checkpoints wherever a candidate or client decision is made.

Outcome

Workload capacity equivalent to two full-time recruiters released. In one measured period an active sourcing pool grew from about 80 to 119 candidates (~49%) while 100+ vacancy signals were reviewed.

Digital Product

iMPLEMENTAi.ae — A Production Website Built With AI Agents

Owner and builder · iMPLEMENTAi.ae

Context

Website design, copy, SEO and releases were slow and expensive to outsource, and prices or offers repeated across pages kept drifting out of sync.

What Was Built

  • The full site — including this page — designed, written and maintained with AI coding agents under David's direction.
  • Offers, prices and WhatsApp messages generated from one source of facts, so every page stays consistent.
  • Page-specific social preview cards with cache-safe image URLs, embedded Stripe checkout and consent-controlled analytics.
  • Releases deployed through a GitHub Actions workflow that checks pages, assets, robots and sitemap before going live.

Outcome

Website, brand, copy and SEO production brought in-house, with changes shipped and verified in minutes.

Enterprise Architecture

AI Estimation Brain (Proposal)

Architect · Major Dubai real-estate developer · 2026

Context

An estimation department held years of quotations and estimates across Excel files, PDFs and scans, but finding a comparable rate meant searching folders by hand.

What Was Designed

  • A source-linked intelligence layer covering documents, cost lines, vendors, items and projects, with hybrid search.
  • Specialist agents for quotation benchmarking, estimate benchmarking, source evidence, indexing and feedback.
  • Every answer linked back to the original file and page, with estimators keeping the final decision.

Status

Architecture and proposal delivered. The design principle: the intelligence layer, not the chat interface, is the asset.

Case Studies

Training & Coaching

Training

Practical Agentic AI Course and David AI Coach

Creator · iMPLEMENTAi.ae · 2026

Context

Most AI courses teach about agents rather than how to build, test and run one. Learners needed a build-first path that works in the AI tools they already use.

What Was Built

  • A free, public 17-module curriculum with lessons, labs, workshops, rubrics, templates and a graduation capstone (GitHub · course page).
  • David AI Coach, an AI tutor that runs the course inside ChatGPT, Claude or Codex using David's method — explain, build, inspect, debug, verify, teach back — and resumes from saved progress.
  • Personalised learner tracks, including a one-to-one technical track built around the learner's trade.

Outcome

A self-paced programme of 50–75 learning hours, currently followed by three learners.

Coaching

Executive Career Method and Hybrid CV

Creator and coach · Talent Bridge · 2021 – Present

Context

Senior professionals were under-represented by their own CVs and profiles, which listed duties instead of the scale and impact recruiters search for.

What Was Built

  • A 7-step Executive Career Method, starting with a deep-dive discovery session and running through executive CV, profile optimisation, job-search strategy and interview coaching.
  • The Hybrid CV format: a one-page career dashboard followed by a detailed career journey.
  • 4- and 8-week Platinum Career Programmes, with the method encoded into AI Skills so quality reproduces consistently.

Outcome

54+ executives, including C-level leaders, coached one-to-one over five years.

Before AI

Foundations

Think Secure LLC · 2008 – 2010

Designed and installed CCTV/IP cameras, access control, electric fencing, gate automation, networks, servers, UPS and satellite systems for commercial clients in Johannesburg.

Emirates Group · 2011 – 2019

Coordinated maintenance, contractors and stakeholders across large corporate facilities using Oracle-based processes — the operating reality AI must fit into.

Hands-On Technology · From 2004

Self-taught PC building, upgrading and trading in Pretoria — the start of twenty years of building and fixing systems.

Next Step
“Understand the real system, improve it, test it against reality — and turn what works into something reusable.”