Johan Steenkamp
Summary
I help organisations adopt AI and carry it through to systems their people use every day. I work across strategy, delivery and the people whose work changes, combining domain expertise, architectural judgement and modern AI tooling.
I start with the business: which process is worth changing, what AI can do for it, and whether an existing tool already does the job. I write production code, so what I recommend is what I can deliver. I work with the people who do the work and the people accountable for it, and turn what they tell me into something engineers and suppliers can build. They are involved from the start, because AI changes what people do all day and adoption fails when that change is not designed for.
Based in Christchurch, New Zealand, I have over a decade of experience in technical leadership and product engineering. I am available for contract, fractional and advisory roles.
Why Judgement Matters Now
AI tooling has made implementation cheap: systems that used to take months now take weeks. The constraint has moved to knowing:
- What to build and what to skip
- How to put AI into a process without losing control of it
- Which architectures work in production
- How to make a complex distributed system something people can explore and understand
Cheaper implementation also creates demand. This is the Jevons paradox: when something becomes cheaper, more of it is used. Bespoke software that was never economically viable for a small firm, or for one process inside a large organisation, now is. As implementation becomes a commodity, domain knowledge and architectural judgement are what remain scarce.
Where AI Works in Production
Three kinds of AI system are viable in production today. They differ in how much the system does on its own and who makes the decision.
- Document intelligence. Processing, classifying and extracting information from unstructured documents at scale: classification that handles thousands of documents, and deep search with observability and evaluation built in. Examples: the Document Substrate, in production use on real client work, and Deep Search, a production research pipeline with evals and telemetry.
- Augmented workflow. AI as one step in an existing process, not a replacement for it. The AI suggests and a person decides. Engineering design assistants, fraud review, operational decision support. Example: the review queue in the Document Substrate. The model suggests a document type and reads the figures, code flags anything it cannot confirm with a plain-language reason, and a reviewer resolves each item and signs.
- Autonomous agents. Systems that complete a task end to end inside set limits, with project-scoped memory, retrieval, tool use, evals and an escalation path to a person. Examples: document processing that routes work on its own classification, applying confident results and holding the rest for a person, and the Cycling Training Intelligence coaching layer.
Project detail and screenshots · Six stages for accounting staff
Principles
- Start from the business, not the technology. Every engagement begins with what the organisation knows and how its people work. The technology follows the problem.
- AI suggests, people decide. The aim is to make your best people's judgement available wherever it is needed, not to automate the people who developed it. People move from doing the work to reviewing it.
- Keep the model out of the parts that have to be defensible. The model reads and suggests. Code checks and decides. A person signs. Personal details are replaced with placeholders before any model call.
- Limit autonomy and test it. An agent runs inside set limits, is tested against real cases before and after every change, and hands over to a person when it is unsure.
- Prototype first, commit second. A working prototype shows whether a build is justified sooner and more cheaply than a specification does. Working code comes before frameworks.
- Build to last, not to depend. Systems keep working and improving after the engagement ends. Your team can maintain and extend them.
- One senior person, end to end. You get me, not a team of juniors, from the first conversation with the business through architecture to production code, working alongside your leaders, technical teams and partners.
The principles behind the Document Substrate
How I Work
I work in three phases, starting from the expertise already in your organisation. Once it is captured and written down, I call it a substrate.
- Map. Short sessions with the people who do the work, starting from the documents they already produce and one real job followed end to end. You receive a written account of how the work is done and an architecture note. Paid for on its own, and yours whether or not we build.
- Architect. The architecture note becomes a design: what already exists, what has to be built for you, and where people stay in control. One install per customer, in accounts you own.
- Build. The parts that are yours, built and run in production. Every correction a reviewer makes is recorded with its reason and becomes a test case, and the system is iterated as real work shows what the first version missed.
Experience
AI Product Engineer
Orbital | Self-employed | Apr 2025 – Present
Building AI systems based on a firm's own expertise, from capturing how the work is done through to production deployment. Built the Document Substrate, a document AI foundation for professional practices, in production use on real client work. Document AI, knowledge capture, evaluation frameworks, MCP servers, React applications and geospatial intelligence.
Principal Frontend Engineer
Darwinium | Full-time | Sep 2022 – Apr 2025
Led development of the fraud detection frontend: UI, dashboards and data visualization for a platform that processes billions of digital interactions. Set engineering practices for modern React development. TypeScript, React, GraphQL, AntD, AntV, Mapbox.
Digital Architect & Head of Software
Syft Technologies | Full-time | May 2018 – Sep 2022
Responsible for end-to-end digital infrastructure and strategy for a scientific instruments company. Led and grew engineering teams. Architected platform systems on AWS, including monitoring for a globally distributed instrument fleet. Active developer in full-stack applications using React, GraphQL and Next.js.
Principal Engineer
Wynyard, Cognevo, & Telstra | Full-time | Aug 2014 – Apr 2018
Principal Engineer and Application Architect across multiple technology companies, focusing on data visualization frontends for security products. Led adoption of GraphQL and modern React patterns. Node.js, React, Redux, GraphQL.
Founder
Orbital | Self-employed | 2000 – Present
At Orbital I led the creation of AssetNow, a content management system that let non-technical users manage and publish web content. I worked with DAS and Infosign in Japan to refine and commercialise it. Released in 2001, when few content management systems were available in Japan, it was adopted quickly, particularly in education and public institutions, and became one of the country's leading content management systems.
Skills
- AI: AI SDK, LLM APIs, Model Context Protocol (MCP), Prompt/Context Engineering, Evaluation Frameworks (Evalite), Observability (Langfuse, OpenTelemetry), Hybrid Retrieval (pgvector, Reciprocal Rank Fusion)
- Document AI: Document parsing and OCR (LlamaParse, LlamaExtract), PII redaction and entity vaults, schema-driven extraction, figures traced to their source page, deterministic verification, durable pipelines (Inngest)
- Knowledge Capture: Completeness specifications, practice manuals, checks that run in code, correction records as test cases
- Geospatial: Multi-source data fusion, temporal-spatial analysis, network graph overlays, interactive map intelligence
- Frontend: React, Next.js, TypeScript, Tailwind CSS, shadcn/ui, GraphQL, Mapbox, Deck.gl, AntD, AntV (G2, G6)
- Backend: Vercel, Node.js, Supabase (database, auth, storage), AWS, DynamoDB, S3, Postgres, SQLite
Education
BSc Eng (Electronics) University of Cape Town