Structured data intake
Inspect Excel or JSON input, identify missing fields, and review proposed repairs before planning.
AGENTIC INITIAL STUDIO
Steel loading.
Multi-agent coordination.
A software and research project connecting data preparation, loading optimisation, agent coordination, and human oversight.
MP-02 / Interactive caseExplore the result.
Explore the result.
Inspect the decisions.
A completed planning run, brought into an interactive dashboard. Explore the regional map, inspect wagon layouts, and compare four strategy agents.
Open interactive case ↗486 products · 18 regions · 81 planned wagons
Synthetic product data. Precomputed results, with no live model calls.
The challengeA loading plan has to do more
A loading plan has to do more
than fit the products.
Steel loading combines competing objectives with strict physical constraints. A useful plan must account for capacity, geometry, destination grouping, and changing operating conditions — and give people enough evidence to review the result.
Capabilities
One connected decision workflow.
Constraint-aware optimisation
Generate feasible plans and check weight, boundaries, overlap, and complete product coverage.
Multi-agent evaluation
An optional council compares strategy-specialised candidates, rejects invalid plans, and ranks valid alternatives.
Human approval
Keep technical validation separate from operational approval, with explicit plan states and versions.
Dynamic replanning
Respond to product changes or wagon availability, with partial and full replanning strategies.
Reviewable evidence
Connect loading layouts, validation summaries, plan history, and audit artifacts to the persisted result.
Architecture
Separate responsibilities.
Keep the evidence connected.
A React interface talks to the Python API. An independent worker executes planning tasks against shared persistent state; model assistance is optional.
Inspect and confirm a canonical input.
Submit a durable, traceable job.
Build and compare candidate plans.
Check the result independently.
Review before activating the plan.
Where LLM and RAG fit
Optional model assistance supports ambiguous input interpretation and evidence-grounded strategy advice. The optimiser and independent validator remain responsible for physical feasibility; a model response does not replace those checks.
Engineering beyond a single run
Durable jobs, versioned plans, artifact integrity checks, and independent worker execution make the result easier to inspect and recover. This page describes the project design, not a claim of production customer deployment.