LLM Shipper Profiles — Adaptive Decision-Rule Generation

End-to-end system that generates literature-grounded crowd-shipper decision rules from traveller profiles, using RAG over a behavioural literature knowledge base + gpt-4o-mini. Direct implementation of the methodology named verbatim in the Aston PhD project description.

GitHub

Highlights

Built before submission to demonstrate the methodology rather than claim it
Repurposes the November 2024 CV-JD Suitability Checker codebase — proves the architecture

Problem

The Aston PhD posting calls for 'use of Large Language Models (LLMs) to generate adaptive user profiles and decision rules' — a methodology with no public reference implementation. Existing crowd-shipping decision models rely on fixed-utility discrete choice, missing the heterogeneity LLMs can capture.

Approach

Pydantic schemas for TravellerProfile, ParcelRequest, DecisionRule
Curated knowledge base of 10 chunks from Punel & Stathopoulos (2017), Archetti et al. (2016), Le et al. (2019), Sampaio et al. (2020)
RAG retriever (sentence-transformers cosine + sklearn TF-IDF fallback)
LLM generator (gpt-4o-mini) producing literature-cited DecisionRule JSON with reasoning
Dual-mode design: deterministic rule-based fallback when no API key, so reviewers can rerun offline

Stack

PythonOpenAI gpt-4o-minisentence-transformersscikit-learn (TF-IDF)PydanticJupyter

Outcome

13 files, 1449 lines, publicly available on GitHub
Demo notebook: 5 walkthrough examples + aggregate threshold distribution
Integration hook into Mini-Digital-Twin agent-based simulation layer