AI PRODUCT DESIGN · ENGINEERING · 2026

Design the behaviour.Build the system.

I turn ambiguous AI capability into operable products: explicit behaviour, inspectable evidence, reviewable actions and production TypeScript.

See the proof Open résumé

Available for senior / staff AI product roles · Cape Town → global

OBJECTSEVIDENCEACTIONSSHIPPED IN CODE

ABOUT / NATHAN

Design is the model of what the product does.

I am a Cape Town-based product designer and engineer. I work between product direction, interaction systems and production code—especially where an AI product has uncertainty, permissions and real-world consequences.

My strongest work starts before the screen: defining objects, states, evidence and action boundaries. Then I move far enough into TypeScript, APIs, data and tests to learn whether the idea survives implementation.

Pangea verifiedTop Earner · $10K+ earned · recommended by clients twice

Base: Cape Town · global teams
Range: Strategy → system → shipped code
Best at: Ambiguous AI + complex B2B workflows

SELECTED WORK

Direction that survives contact with the system.

Four evidence-bounded cases: shipped systems, public methods and portfolio-safe visual artifacts. No invented screens or unsupported outcome metrics.

01 / CABOO

From a document to a grounded agent — and then a safe action.

Overview I defined and built Snap, a product that compiles business documents into source-grounded AI Passports while preserving the evidence needed to trust and act on them.

Role: Founder & Design Engineer · product model, interaction design and production implementation
Problem context: Ordinary AI context loses provenance, unknowns, tables and the boundary between a request and a completed real-world action.
Solution: A TypeScript, Hono and Postgres document compiler, knowledge graph, provider handoffs and MCP-capable discovery/action gateway.
Impact / evidence: A working public product with current source, boundary, route, promise and fixture gates passing. Commercial outcomes are deliberately omitted until retained analytics exist.

View the live product Open the proof

A hand-drawn production line turning a document into a source-grounded agent and an action.

END-TO-END CASE / RESEARCH → MEASUREMENT

I treated trust as a product system, not a disclaimer.

This is the full arc behind the shipped product: the inquiry, the product model, the decisions that survived implementation, and the quality signals I used after delivery. The measurements below are product-evaluation results, not customer or revenue claims.

01 / RESEARCH
Separate durable signals from AI-search hype.
I directed a research brief across discovery, crawler behaviour, structured data, MCP, agent authentication and booking. It exposed the gap: assistants could read fragments, but reliable action still needed accurate first-party evidence and an explicit control boundary.
02 / CONCEPT
Reframe the object from “document upload” to “evidence passport.”
I defined the core promise around facts, provenance, unknowns and visual evidence that remain attached as content moves from a source document into an AI conversation.
03 / SYSTEM
Model the states before polishing screens.
I mapped the creator and scanner journeys, provider handoffs, review states and the action boundary: draft or request → human review → verified external outcome. That model became the interface and API contract.
04 / DELIVERY
Build far enough to test the product truth.
I implemented the compiler, knowledge graph, provider profiles, public resources, action review and MCP-capable gateway in TypeScript, Hono and Postgres—then shipped the public Snap surface.
05 / MEASUREMENT
Measure grounded behaviour and transport integrity.
I evaluated whether the delivered system preserved evidence, remembered a stated priority, surfaced unknowns, resisted false premises and kept an enquiry explicitly unsent until review.

25 / 25behaviour checks passed

A logged-in, ten-turn Claude acceptance session covered grounding, visual evidence, unknown recognition, memory, truth scope, contact use and action honesty.

163 / 163artifact-contract checks passed

The current ChatGPT and Claude public build artifacts each passed the deterministic contract audit for prompt, source, action and handoff integrity.

A hand-drawn product model showing state and downstream consequences outside a screen.

02 / BEROE

Object-Oriented Design for agentic enterprise workflows.

Overview I created a method for turning probabilistic agent behaviour into explicit product objects, conditions and implementation consequences.

Role: Senior Product Designer · agentic procurement and enterprise workflow design
Problem context: Confidence, uncertainty, review, partial completion, recovery and handoff could not be understood as one happy-path conversation.
Solution: An object registry, condition grids, two-hop change-impact analysis and two tiers of agent-executable handoff.
Impact / evidence: The method makes behaviour discussable across product, design and engineering. Client screens and quantitative outcomes remain confidential or withheld until their evidence is retained.

Honest limitation This public case shows the operating method, not a fictional recreation of the confidential product.

03 / MTN

One reusable foundation for products moving in different directions.

Overview I led design-system creation and governance across a multi-surface telecom product ecosystem.

Role: Senior Product Design Manager · Nov 2021–Nov 2024
Problem context: Multiple product verticals needed shared primitives without forcing unlike workflows into one rigid template.
Solution: Reusable components, token architecture, governance and common design-to-engineering practices paired with cross-app redesign work—including adaptive dark and light directions for MTN eShop.
Impact / evidence: Established reusable foundations and shared delivery practices across multiple verticals. The eShop frame is a public portfolio artifact; no unretained performance percentage or team-size claim is attached.

System foundationsGovernanceProduct direction

Dark-mode MTN eShop concept showing navigation, service categories and a 5G router promotion. — PUBLIC ARTIFACT + SYSTEM MODEL The interface shows a direction; the illustration explains the operating system behind it.

A deadpan Xiaohei character operates a hand-drawn loom that converts scattered product fragments into reusable foundations. — PUBLIC ARTIFACT + SYSTEM MODEL The interface shows a direction; the illustration explains the operating system behind it.

Publisher Prime interactive helicopter training experience with a cockpit, collective control and instructional prompt. — INTERACTIVE EBOOK ARTIFACT A flight principle becomes something the learner can manipulate, observe and understand.

04 / PUBLISHER PRIME

Turn the lesson into a cockpit, not another page of instructions.

Overview I helped shape an interactive eBook for helicopter-pilot training—blending storytelling, explanation and direct manipulation so students could explore flight principles rather than only read about them.

Role: Product design · interactive learning experience
Problem context: Technical flight principles are difficult to absorb when video, text, instruction and physical control remain separate.
Solution: A simulated learning environment combining video, explanatory text, game-like controls and in-context instruction.
Impact / evidence: The public artifact demonstrates the interaction direction and learning model. No completion, retention or training-performance metric is claimed without retained evidence.

Interaction designLearning systemsVisual storytelling

Open the seven-page proof

Public-safe proof from current product, Pencil and Pangea-owned portfolio material.

WORKING PRINCIPLES

The rules behind the cases.

A hand-drawn Xiaohei character building a bridge toward uncertainty, review and recovery states.

01 / BEHAVIOUR

AI product design starts where the happy path ends.

I design uncertainty, review, partial completion, recovery and handoff as product behaviour — not as edge-case copy added after the interface is drawn.

A hand-drawn Xiaohei character operating a machine that prints a receipt with source, provenance and unknown markers.

02 / EVIDENCE

Every answer should carry its receipts.

Sources stay attached, provenance remains inspectable and unknowns stay visible. Trust comes from evidence a person can examine, not confidence styling.

A hand-drawn action system separating a request, review and confirmed external outcome.

03 / ACTION BOUNDARY

A tool call is not a completed action.

The product may draft and request. Only an external confirmation earns the word done. That boundary protects users without making useful action impossible.

CONTACT / CAPE TOWN

Give me the ambiguous AI problem.I’ll make it operable.

I am exploring AI Product Design Engineer, Design Engineer and Staff / Principal Product Designer roles with global teams.

Start a conversation LinkedIn