Client

Global commodities and equities trader.

Challenges

- The traders relied on personal and company wide spreadsheets / balance sheets, informal chat channels, public chat forums, public and private data sources.

- The datasets are unstructured, sometimes messy and fragmented.

- There are inconsistent data standards across regions, including languages, units, estimates v actuals, unreliable risk adjusted datasets, an ability to extract numbers or risk patterns from words.

- Some of the data requires heavy processing, e.g. weather, yields, NASA.

- Decisions need to make allowances for structural breaks such as wars, sanctions, tariffs, extreme weather etc.

- Multiple systems and databases, manual spreadsheets, silos.

- People missing data and making biased, emotive and ill-informed trading decisions.

- Protecting IP, confidential information, trade secrets and formulas within the client’s balance sheets.

- Building a clean, real-time, unified data layer can be resource intensive.

Solution

Zai Node is delivering a hybrid solution which is a human-led strategy with AI-augmented decision support and automated operational optimisation.

The first major data fragmentation challenge is solved by our engineers and data scientists building a centralised data lake which auto pipes real time and manual data (e.g. markets, shipping, weather, crop yields, stock reports, credit).

Using a VLM and LLM to extract data from unstructured sources incluing satellite imagery and public chat forums in many languages.

We have standardised and cleaned the data and created a “single source of truth”. We think of this as a “digital twin” of a commodities trading balance sheet.

Using ML model to help detect structural breaks from extreme events and stress-test the client’s models under historical crises, looking for probability and anomaly detection.

Impact

Data pipeline solution improves decision making speed and realiability, and eliminates spreadsheet risk. We’re predicting a 1–3% improvement in trading margins from speed and less errors, and a 30% reduction in operational time and resources.

Institutional knowledge is captured instead of lost. It becomes codified, measurable, repeatable and scalable, moving from a individual balance sheet approach.

Upshot

- Clean, structured, automated data pipe.

- Use AI to augment trader judgment.

- Automate operations and data processing.

- Use AI to help collect and translate data, detect risk, improve probability.

Client

State government agency responsible for transport, utlities, civic assets.

Challenges

- Fragmented data in silos and paper records.

- No unified real-time visibility across agencies across asset classes, leading to reactive maintenance, breakages, high emergency costs and service disruption.

- Disconnected capital planning, duplication of works, public scrutiny, competing funding priorities, and a need for transparency.

- Infrastructure stress and utilisation was not modelled dynamically.

- Security and sovereignty concerns with expanding attack points.

Solutions

Our team’s involvement focused on the centralised, standardised asset data layer integrating BIM, GIS, IoT, sensor data, SCADA and various operational systems.

We also provided advice on real-time monitoring and predictive analytics, model selection and tuning, and sensor monitoring with predictive maintenance modelling.

Impact

- Single source of truth across infrastructure portfolio.

- Reduced reactive maintenance and service disruptions.

- Improved emergency response coordination.

- Evidence-based capital allocation.

- Enhanced transparency to government and the public.

- Reduced cyber and operational risk.

Client

National quick-service restaurant chain

Challenges

- Business improvement project revealed considerable costs, cycle times, delays, errors, rework and value loss in the procurement process.

- There was often uncertainty about what the business rules were, which person can make decisions about purchase categories, what the company has previously paid (and agreed terms) for like categories, and the contracting process.

- Creating RFIs, RFPs, assessing bids and commercials was manual and often didn’t refer to historical data.

- Ave 132 total labour hrs per contract until commencement of deliverables, and 11 weeks cycle time.

Solution

_ Current state analysis to better understand the problem, root causes, goals, and relevant data and systems.

- Clean, structure and optimise existing datasets.

- Use a public model to assist draft RFI/RFP templates, tailor questions by category, summarise vendor responses and highlight non-compliance.

- For bid responses we used a VLM to extract pricing, delivery timelines, and compliance statements from vendor submissions automatically.

- To assess the bids we used an LLM based classification model trained on historical award data, and then score suppliers on cost, risk, quality, and other defined metrics.

- Agentic AI tool analyses vendor documents for risk and compliance issues (e.g., sanctions, ESG policies, cyber posture, AML, Chain of Custody etc.) using Hugging Face NER pipelines and GPT-based RAG over vendor data.

- Internal approvals used Power Automate, agentic AI and MCP servers to process mine and task automate. Identify bottlenecks, predict approval delays and route workflows dynamically.

- Predict spend by category, supplier, or region, detect anomalies or non-compliant spend, using classical statistical approaches combined with fine-tuned LLM for post-analysis and Power BI.

- Finally, creating a Procurement Assistant chatbot, RAG with text embeddings and our custom RAG fine-tuned LLM. E.g. “Show top suppliers by A, B and C.”

Impact

70–80% reduction in manual labour from “we need to buy” until contract delivery start.

55% faster cycle times from RFI to contract signed.

Full audit trail, risk visibility and decision consistency.

Accurate single source procurement spend forecast across whole group.

Labour average is 45hrs per contract (down from >130).

All FAQs are handled online by Q&A chat with only 12% going to human.