The chatbot replaced a USD5,000/month vendor contract. The more significant change is operational: marketing and technical support now share responsibility for keeping the knowledge base current. Marketing owns product announcements, campaign context, and pricing updates; support owns integration guidance, error resolution, and account queries. Two teams now maintain an AI system together as a default part of their workflow.

The most technically complex system in the portfolio. Five interconnected workstreams replacing manual agency processes across Meta, Google Ads, and TikTok, built on a self-hosted production n8n instance. The system runs autonomous campaign setup, optimisation, and reporting at the infrastructure level the team would otherwise outsource to agencies.

Built to establish objective performance measurement across APAC, tracking three media agencies that run paid digital across Meta and Google Ads. The system creates a data-normalised accountability layer between the client and its agencies, replacing subjective slide deck reviews with repeatable AI analysis.

The real change is behavioral. Marketing teams now walk into agency review meetings with an AI meeting assistant active on screen. They query it in real time: what drove the performance delta, which campaigns underperformed against forecast, what to push back on. The agency prepares differently because the team arrives better informed. The dynamic in the room has changed.

This pipeline produces the high-volume, content-team-owned work that sits alongside subject-matter expert authoring: guides, educational explainers, vertical landing pages, localised market variants. It does not replace the SME-authored thought leadership that lives elsewhere in the content function. It handles the templated production work that otherwise gets outsourced to agencies or deprioritised due to capacity.

The same architecture runs trilingual at PURE (EN, TC, SC) and replaced two content agencies. For a global business with regional content needs, the unit of work is not one article: it is one canonical piece plus its market variants, social derivatives, SEO metadata, and structured data schemas. The pipeline produces all of them from a single brief.

A standalone application built with React and deployed on Cloudflare Workers and Pages, with Cloudflare R2 for image storage. Accepts a reference image, uses Claude API to describe the environment, then calls FAL.ai to generate AI images based on that environment description. Four visual types are supported: promotional with customisable offer text, UGC-style, community-based, and editorial ghost showing a person standing alongside a ghost-like figure representing an aspirational identity. All generated assets are resized into 9 formats covering Google PMAX, Meta, and Programmatic Display specifications. Logo compositing runs in-browser using Canvas API with three positioning options. Text overlays are fully customisable before export with font and size preserved across edits.

The underlying architecture, prompt-driven creative generation with brand constraint inputs, is format-agnostic. The same system is being extended to generate organic content templates for the social media team, who currently produce content manually. One build, multiple use cases, expanding adoption.

A market-agnostic review-response architecture, currently in production across two of the client's markets but designed so adding a new market is a configuration file, not a code change. Locations, Telegram approval channel, brand voice, and review-scoring thresholds are all per-market parameters. The same workflow scales to as many markets as a business operates in.

Replaced a manual, inconsistent process where reviews were answered sporadically. Every review now receives a contextual, brand-appropriate response with a human final check before anything is published.

Regional Instagram and Facebook performance analysis covering both markets. Each region pairs a scheduled analysis workflow with a conversational chat assistant, turning one-directional reports into interactive intelligence tools the social team can query in plain language.

The team was already doing ad hoc AI analysis by manually copying Instagram data into external LLMs. What changed was the setup cost: zero. All data is structured, embedded, and ready. The AI chat is open during weekly performance reviews. The team asks it for content recommendations, underperformance diagnoses, and posting schedule optimizations without leaving the dashboard. The behavior shifted because the friction disappeared.