Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Project Synapse: A Hierarchical Multi-Agent Framework with Hybrid Memory for Autonomous Resolution of Last-Mile Delivery Disruptions

Arin Gopalan Yadav, Varad Dherange, Kumar Shivam

TL;DR

Project Synapse tackles last-mile delivery disruptions by proposing a hierarchical multi-agent framework powered by a hybrid memory architecture and an end-to-end workflow orchestrated with LangGraph. The system employs a Resolution Supervisor to decompose complex disruption cases and delegates execution to specialized worker agents, all grounded by a retrieval-augmented memory system and secured through the MCP Toolkit. Evaluation on a benchmark of 30 realistic disruption scenarios derived from thousands of reviews uses an LLM-as-a-Judge protocol with bias mitigation, reporting strong performance, especially in reasoning quality ($0.77$) and overall effectiveness ($0.73$). The results validate the viability of autonomous disruption resolution in dynamic logistics, while acknowledging limitations due to scale and simulated deployment, and suggest directions for self-evolving hierarchies, RL-based planning, and multimodal extensions. The work has broader implications for adaptive, memory-grounded agent systems across domains such as logistics, customer service, and emergency operations.

Abstract

This paper introduces Project Synapse, a novel agentic framework designed for the autonomous resolution of last-mile delivery disruptions. Synapse employs a hierarchical multi-agent architecture in which a central Resolution Supervisor agent performs strategic task decomposition and delegates subtasks to specialized worker agents responsible for tactical execution. The system is orchestrated using LangGraph to manage complex and cyclical workflows. To validate the framework, a benchmark dataset of 30 complex disruption scenarios was curated from a qualitative analysis of over 6,000 real-world user reviews. System performance is evaluated using an LLM-as-a-Judge protocol with explicit bias mitigation.

Project Synapse: A Hierarchical Multi-Agent Framework with Hybrid Memory for Autonomous Resolution of Last-Mile Delivery Disruptions

TL;DR

Project Synapse tackles last-mile delivery disruptions by proposing a hierarchical multi-agent framework powered by a hybrid memory architecture and an end-to-end workflow orchestrated with LangGraph. The system employs a Resolution Supervisor to decompose complex disruption cases and delegates execution to specialized worker agents, all grounded by a retrieval-augmented memory system and secured through the MCP Toolkit. Evaluation on a benchmark of 30 realistic disruption scenarios derived from thousands of reviews uses an LLM-as-a-Judge protocol with bias mitigation, reporting strong performance, especially in reasoning quality () and overall effectiveness (). The results validate the viability of autonomous disruption resolution in dynamic logistics, while acknowledging limitations due to scale and simulated deployment, and suggest directions for self-evolving hierarchies, RL-based planning, and multimodal extensions. The work has broader implications for adaptive, memory-grounded agent systems across domains such as logistics, customer service, and emergency operations.

Abstract

This paper introduces Project Synapse, a novel agentic framework designed for the autonomous resolution of last-mile delivery disruptions. Synapse employs a hierarchical multi-agent architecture in which a central Resolution Supervisor agent performs strategic task decomposition and delegates subtasks to specialized worker agents responsible for tactical execution. The system is orchestrated using LangGraph to manage complex and cyclical workflows. To validate the framework, a benchmark dataset of 30 complex disruption scenarios was curated from a qualitative analysis of over 6,000 real-world user reviews. System performance is evaluated using an LLM-as-a-Judge protocol with explicit bias mitigation.
Paper Structure (49 sections, 5 theorems, 15 equations, 5 figures, 5 tables, 2 algorithms)

This paper contains 49 sections, 5 theorems, 15 equations, 5 figures, 5 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. The Last-Mile Challenge in the Super-App Era
  3. The Disruption Problem
  4. The Agentic Systems Paradigm Shift
  5. Introducing Project Synapse
  6. Related Work
  7. The Evolution of LLM-based Agents
  8. Hierarchical Multi-Agent Systems for Complex Task Decomposition
  9. Cognitive Architectures for Agents: Planning, Tool Use, and Memory
  10. Planning and Tool Use
  11. Memory Systems
  12. AI in Logistics and Supply Chain Disruption Management
  13. Evaluation Paradigms for Generative Agents
  14. The Synapse Framework: System Architecture and Methodology
  15. Mathematical Preliminaries and Notation
  16. ...and 34 more sections

Key Result

Proposition 3.5

A complete execution of the Synapse system on event $e \in \mathcal{E}$ produces a trace: where $T \in \mathbb{N}$ is the termination time, and $v_T \in V_{\text{terminal}}$ is a terminal node producing either a resolution report or an escalation signal.

Figures (5)

  • Figure 1: Overview of the Project Synapse Hierarchical Multi-Agent Framework and Resolution Workflow.
  • Figure 2: Detailed architecture of the Project Synapse system showing the interaction between components.
  • Figure 3: The linear projection derived from PCA captures the dominant global trends within the dataset. Separation of clusters along the principal components reflects variations in overarching thematic dimensions, such as payment-related issues, delivery performance, or driver behavior.
  • Figure 4: We conducted a structured analysis of 6,239 user reviews collected from Google Play Store and Apple App Store across the superapp ecosystem applications.
  • Figure 5: The cosine-similarity heatmap of TF-IDF embeddings offers a global overview of inter-scenario relationships, allowing for the detection of semantically coherent clusters that correspond to recurring service-failure themes.

Theorems & Definitions (15)

  • Definition 3.3: Directed Conditional Graph
  • Definition 3.4: State Transition Dynamics
  • Proposition 3.5: Execution Trace
  • Remark 3.6
  • Definition 3.7: Hybrid Memory System
  • Definition 3.8: Semantic Retrieval
  • Definition 3.9: Augmented Generation
  • Theorem 3.10: RAG Consistency
  • proof
  • Proposition 3.11: Time Complexity of Single Resolution
  • ...and 5 more