Work that turns messy operations into intelligent systems.

Helped the team move past a narrow predictive-maintenance brief and into a multi-agent operating model where specialised agents — anomaly detection, root-cause analysis, QA, explanation — collaborate through orchestrated workflows. Beautec's contribution was the commercial framing, the architecture call, and the risk controls around it.

Verification agents cross-check outputs, structured retrieval anchors responses to ground truth, and confidence thresholds gate downstream automation. Agents are far more useful when they QA each other than when they operate solo.

Prototype work — designed to test commercial value quickly before larger investment.

Turned an exploration bottleneck into an executive-level opportunity: less manual review pressure while keeping geologists in control of the final decision. Earlier attempts treated drill-core analysis as a classical ML problem and stalled on image consistency. Modern vision agents collapsed much of that complexity.

A reference-led architecture: classified samples anchor the agent, which compares unread cores against known examples. The output is a prioritised review queue for human experts — not an automated geological decision.

Prototype work — designed to test commercial value quickly before larger investment.

Helped scope a drive-through personalisation system that connects customer experience, operational efficiency and a measurable commercial number. The architecture separates responsibilities cleanly: a vision agent identifies the vehicle, a retrieval agent fetches order history, a recommendation agent generates the upsell, and the menu adapts dynamically.

Modelled upside: +10–15% AOV, ~20% waste reduction, 5–10% retention lift — with structured handoffs keeping each decision verifiable.

Prototype work — designed to test commercial value quickly before larger investment.

Took the drive-through agent architecture and tested it in a second vertical — a café operator — to see what was genuinely reusable and what needed re-tuning. The core architecture (identity, retrieval, recommendation, dynamic presentation) held cleanly. The menu, product mix and operating model flexed to the new context.

The output: a clear view of which parts of the solution were portable, which needed vertical-specific work, and how the build could become a repeatable service offering rather than a bespoke prototype.

Prototype work — designed to test commercial value quickly before larger investment.

Built a cross-industry inspection prototype to show whether vision-plus-LLM agents could support safety, asset condition and infrastructure monitoring. The goal wasn't image classification; it was explainable reasoning a supervisor, engineer or safety lead could act on.

One test used thermal imagery of solar panels: the agent reasoned that yellow and red zones meant overheating cells while uniform blue and purple suggested healthy operation. The capability isn't the vision; it's the reasoning.

Prototype work — designed to test commercial value quickly before larger investment.

Worked through a design-stage engagement for a citizen-facing council-services assistant. The hard part wasn't whether GenAI could answer the questions — it was whether the organisation could safely expose it to the public. The work centred on a custom AI risk governance framework — Fair, Accountable, Intelligent, Responsible, Aligned.

Made public-sector accountability requirements explicit and architecturally enforceable before deployment: auditability, escalation paths, role-based access, PII controls.

Scoped an early-stage POC for LLM-assisted assessment that could reduce marking effort without weakening educational quality or accountability. The model scored student answers against multi-dimensional rubrics — accuracy, completeness, understanding, clarity — with the human assessor staying accountable for the final judgement.

Defined the safe operating boundary: where AI could accelerate, where human review was required, and how rubric-based scoring could be made explainable enough for educators to trust.

Prototype work — designed to test commercial value quickly before larger investment.

Strategy and design engagements across data foundations — turning executive intent into platform direction and investment pathways. Modern data platform architecture for an Australian port operator with a lakehouse design positioned to support downstream ML use cases.

Data-as-a-Service for a renewables generator, optimisation algorithms for an energy operator's load and dispatch, and a data strategy for a national transport operator covering demand forecasting, route optimisation and fleet predictive maintenance.

Sizing addressable demand for an established golf equipment business, mapping competitor positioning, and pointing to the highest-leverage growth moves. Currently in active engagement.

Growth strategy for a brand moving from regional specialty retailer to national premium presence. Pitch and engagement plan built; outreach in progress.