We take complex ideas, products, and research and reduce them to the version you can hold in one hand — using observations, tools that already exist, and principles that already work. Shared openly through the newsletter and advisory work.
02 — Build
Then we ship it.
Clear thinking is only half the job. We build thoughtful products and services with the best current tools and the timeless principles of product, design, and technology — with honesty to the craft.
Clear thinking, built.
How I work
The frameworks driving repeatable delivery.
Six principles that run underneath every system — from the first brief to the go-live checklist.
01
Scope before speed.
Translating ambiguous business problems into concrete, scoped systems with a strict definition of done.
02
Connectors, not copies.
Separating invariant AI reasoning from variant platform APIs to prevent rebuilds.
03
Design for adoption.
Optimising for end-user workflows and business integration, rather than pure technical elegance.
04
Validatable delivery.
Shipping in testable phases. Eliminating the risk of big-bang releases.
05
Documented context.
Codifying architectural trade-offs so the broader team can operate autonomously.
06
Post-launch telemetry.
Evaluating product success through quantitative reranking and usage data, not assumptions.
Selected work
Enterprise AI, scoped and shipped.
A sample of systems taken from a one-line brief to production — across commerce, documents, search, and internal tooling.
AI-native commerce across four enterprise backends
One conversational AI core. Four commerce platforms. Zero custom integration per deployment.
Agentic AI · Commerce · 2025 →
The problem
Enterprise sales teams couldn't resolve product queries, pricing, or availability in real time. Every conversation ended with a delay.
What I built
A conversational layer: chat UI → AI gateway → live catalogue retrieval → transaction routing across Salesforce, SAP Hybris, Shopify, and Shopware — without rebuilding the AI stack per platform.
Outcome
5+ B2B leads generated. Natural language purchasing live across four enterprise contexts. Demoed for four businesses in six weeks.
VLM document verification, in production in two weeks
Replaced a three-day manual review cycle with a vision-model pipeline that automates 90% of document processing.
VLM · GPT-4o · 2025 →
The problem
Manual review of 100+ year-old enterprise supplier certificates was slow, error-prone, and risked supply-chain disruption.
What I built
A VLM-powered ingestion pipeline: webhook intake → GPT-4o vision classification and field extraction → structured validation with confidence scoring. On-prem deployment with Hybris integration.
Outcome
Production-grade deployment in two weeks. 90% of document processing automated. Client validated and ready to scale.
Stack · Webhooks · GPT-4o Vision · Hybris · Multi-format support · Field extraction · Human-review flagging
Replacing brittle ETL scripts with a no-code data extraction layer
Non-technical teams connect enterprise sources and run extraction pipelines independently — no engineer required.
Connectors · No-code · AWS · 2025 →
The problem
Enterprises needed usable data out of SAP, Salesforce, and Hybris — but extraction relied on slow, brittle, IT-heavy scripts. Siloed integrations delayed analytics and AI use cases.
What I built
A universal connector framework with a schema-normalised extraction engine, plus a no-code visual workflow builder that non-technical teams can operate independently.
Outcome
5 leads generated. Partnership with a leading process intelligence company. No engineer required to connect a new enterprise system to the extraction pipeline.
Stack · JavaScript · Node.js · AWS · SAP P2P · SAP O2C · Reusable connector-first architecture
A RAG reasoning engine that makes CRM data accessible in Microsoft Teams
CRM data was rich but inaccessible. Built a Teams bot that surfaces retrieval-grounded answers in seconds.
RAG · Azure AI Search · Teams · 2025 →
The problem
CRM data was rich but inaccessible. Analysts ran manual queries; account teams couldn't surface insights during live calls.
What I built
A Teams serverless bot backed by a RAG reasoning engine — querying Azure AI Search over a structured data lake, returning retrieval-grounded answers in seconds via a prompt library.
Outcome
Technically robust pipeline, client-delivered and validated. Reinforced a strategic filter: build only for product-led companies where architectural decisions compound.
Stack · Azure ADLS · Azure AI Search · Azure OpenAI · Teams Serverless Bot · RAG
A hybrid DeepAgent for automated lead capture
An agentic workflow that turns inbound enquiries into structured, actionable data — no manual triage.
Agentic AI · DeepAgents · 2025 →
The problem
An Australian scrap-metal recycler was losing leads to manual triage — inbound enquiries sat unprocessed and analytics were non-existent.
What I built
A hybrid DeepAgent that intercepts inbound enquiries, classifies intent, extracts structured lead data, and feeds analytics — operating autonomously without human triage.
Outcome
Automated lead capture and analytics pipeline. Inbound enquiries converted to structured, actionable data at the point of contact.
Stack · DeepAgents · Agentic workflow · Structured data extraction · Lead analytics
Writing
Short, honest notes on building simpler things.
The newsletter
One short note every other week. No noise, no filler, nothing to scroll past.
The instinct everywhere is to add — more features, more steps, more noise. We do the harder work of taking things away until only what matters is left.
We think, distilling complex problems into something a person can actually hold; and we build, shipping thoughtful products with current tools and timeless principles. No theatre, no manufactured complexity.