I am a computer science engineer working in supply chain delivery, which means the AI on this page was built by the person accountable for the go-live. That combination is the whole argument.
Retrieval pipelines over incident repositories, configuration documentation and planning master data — chunking, embedding and re-ranking tuned so answers cite the source record rather than inventing one. Built against live delivery-programme data, where a wrong answer costs a cutover.
An agent trained on the historical incident repository that triages recurring issues and suggests resolutions — cutting P3/P4 resolution time by 14% and replacing manual control-tower lookup.
Model Context Protocol servers that connect language models to business systems and internal data over a typed tool surface, so the model reads real state instead of a stale export.
A chatbot covering tool functionality, configuration and ACC-versus-PROD deltas, cutting repeated back-and-forth during implementation and cutover. Plus LLMs applied to scenario modelling, decision validation and requirement drafting inside live programmes.
AI-assisted applications shipped whole — front-end, back-end and API integration — including an ERP and accounting web application and a combined SEO and CRM tool.
In a planning programme a confidently wrong answer is worse than no answer. Every retrieval pipeline I build cites the source record, and the failure mode is "I could not find it" rather than a fluent invention.
The triage agent learns from the actual historical incident repository — years of real defects, real resolutions, real language the planners use. Generic models do not know what ACC-versus-PROD means on your programme.
A 14% cut in P3/P4 resolution time is the point. If a model does not move a delivery metric, it is a demo. AI work gets held to the same reporting line as everything else on the plan.
Front-end, back-end, API integration. I have shipped an ERP and accounting web application and a combined SEO and CRM tool the same way — the model is one component in a working system, not the whole deliverable.
Every shift below is what the market says is coming next. Against each is what I have already delivered, because a claim about the future is only worth reading if it is backed by something that already shipped.
The shift
Through 2026 AI in supply chain moves from dashboards and recommendations to agents that identify risks, propose workarounds and trigger corrective action inside trusted guardrails.
Already delivered
I already run one. The incident triage agent works on the live repository of a delivery programme, proposes resolutions and cut P3/P4 resolution time by 14%. It is not a pilot.
The shift
As agents act rather than advise, the scarce skill becomes setting guardrails, auditing outputs and deciding where autonomy stops.
Already delivered
Every pipeline I build is designed to abstain rather than guess, and to cite the source record. In a cutover, a confidently wrong answer is worse than no answer — that constraint is enforced in the system, not the prompt.
The shift
Copilots embedded in planning workspaces absorb repetitive analysis so planners focus on scenario choice, exception management and stakeholder communication.
Already delivered
The configuration and ACC-versus-PROD chatbot exists precisely to remove the lookup tax from senior people during cutover, so their time goes to the decisions only they can make.
The shift
Employers report they cannot find people who combine deep planning experience with real AI fluency, and note that candidates certified on an APS *and* having led enterprise implementations are a small population.
Already delivered
OMP and Kinaxis certified, 5+ go-lives delivered as rollout lead, and a computer science degree I use daily. That is both halves, not one and an interest in the other.
Client and business names are withheld. What matters here is the shape of the problem, the stack, and why it maps onto supply chain and planning work.
Agri-export operation · organic spices, superfoods, essential oils
A single operational surface for an export business: demand and marketplace signals pulled from search, advertising and marketplace APIs on a schedule, normalised into one database, then scored and turned into prioritised actions by an LLM runtime. Includes automated reporting, anomaly alerting and a task pipeline that assigns work rather than just charting it.
Why it is relevant
The same pattern as a supply chain control tower — multi-source ingestion, exception scoring, prioritised action — applied to the commercial side of an export operation.
Small and medium manufacturers and traders
Double-entry accounting with statutory tax compliance, built around a natural-language query engine so a business owner can ask about their own ledger instead of navigating report screens. Conversational interface over a messaging channel for capture and lookup away from a desk.
Why it is relevant
End-to-end ownership of a transactional business system — the same master data, document and compliance discipline that ERP integration work demands.
Accounting ERP → analytical database
Command-line pipeline that extracts ledgers, inventory and voucher data out of a widely used accounting ERP and lands it in a database server for downstream reporting and analysis, with incremental loads and schema mapping.
Why it is relevant
Master data extraction and mapping between a system of record and an analytical layer — the unglamorous work that determines whether an APS implementation has clean inputs.
Regulated financial planning
A planning tool whose recommendations are constrained by a regulator’s rules, with every output traceable to a cited source. Built so the model cannot produce advice the regulations do not permit — the constraint is enforced in the system, not in the prompt.
Why it is relevant
Grounded, constrained generation in a domain where a wrong answer has consequences — the same design discipline behind AI inside a live planning programme.
The early ones are embedded systems from the Innovation Cell. The recent ones are retrieval pipelines and agents inside Fortune 500 delivery programmes. Same instinct, eight years apart.
Agent trained on the historical incident repository; triages recurring issues and proposes resolutions. Cut P3/P4 resolution time by 14%.
Covers tool functionality, configuration and ACC-versus-PROD deltas, compressing back-and-forth during implementation and cutover.
Full-stack AI-assisted build — front-end, back-end and API integration.
Combined SEO and CRM application with automated data pipelines and agent-driven reporting.
Integrated and optimised Google products to support global COVID-19 vaccine distribution.
Digitised procurement-through-shipping workflows in Excel-VBA and SQL, plus a real-time supplier-risk dashboard.
Prototype flash-flood warning system for local picnic spots, detecting rising water in real time to alert visitors.
Controlled home appliances through a self-designed mobile application over the internet.
Automated crop watering for rural farmers, reducing manual intervention, time and effort.