A Java harness that keeps the LLM away from the wallet

What it does

Delivery Harness is a Java/Spring Boot MVP that demonstrates how to integrate an LLM into on-demand delivery operations without handing it the keys. It runs two synchronous workflows—abnormal-order analysis and compensation suggestion—using synthetic business tools, a deterministic order timeline, and an in-memory knowledge base. Apart from the configured model endpoint, the default application has no required external services and runs fully offline.

The interesting bit

The architecture deliberately strips the LLM of authority: a rule engine alone decides compensation amounts, while the model merely formats reasoning and flags cases for human review. The evaluation scorers even report “not measured” when a case declares no expectation, which is a refreshingly honest approach to benchmarking an MVP.

Key highlights

• Two fixed workflows with deterministic baselines the model must beat, plus synthetic order scenarios that vary by ID
• Four in-memory business tools (orders, ETA, station capacity, compensation rules) that derive results from arguments, making the pipeline testable without real integrations
• Compensation amounts are owned by the rule engine, not the LLM; guardrails enforce human-review thresholds
• OpenAI-compatible client interface lets the 112 tests run against stubs—no Ollama or Docker daemon required for CI
• In-memory lexical retrieval, trace storage, feedback, and evaluation scaffolding with explicit capacity limits
• Degraded-mode reporting when optional dependencies fail, keeping deterministic results available while recording the fault

Caveats

• Everything is synthetic and in-memory: no durable persistence, no production system integrations, and no authentication or rate limiting
• Retrieval is lexical term overlap, not semantic search; CJK text is split on whitespace and punctuation rather than segmented, so recall depends on shared phrases
• The six seeded evaluation cases and lexical scorers are calibrated to catch regressions, not to certify real-world accuracy
• The authors explicitly warn this is an educational MVP and should not be exposed to the public internet

Verdict

Worth a look if you are building guardrails around LLMs in logistics or financial workflows and need a concrete Java reference for separating model output from business logic. Skip it if you want a production-ready delivery platform or a vector-based RAG system.