How Monus AI Cuts Search Costs by 34% with Amazon Bedrock AgentCore

Product Overview

Monus AI, launched by Nanjing Altre Technology, is an AI‑driven e‑commerce search application that focuses on consumer decision support. Its core functions include specification‑level price comparison, fake ad detection, product comparison and intelligent dialogue, aiming to provide efficient and trustworthy decision assistance before purchase.

Challenges in E‑commerce AI Search

The article identifies three major challenges: (1) recognizing the user's decision‑making stage across multimodal inputs (text, voice, image); (2) handling heterogeneous product specifications across platforms, leading to “same‑meaning, different‑name” issues that hinder real‑time price comparison; (3) insufficient matching between user profiles and products due to limited reasoning capability of traditional recommendation systems.

Solution Architecture

Monus AI adopts a self‑developed multimodal fusion input technique to process text, voice and image simultaneously. It introduces a “consumer decision‑stage judgment” mechanism that analyzes semantic features and sentiment, achieving 94% accuracy in identifying whether a user is in the demand‑germination, information‑gathering or purchase‑decision phase.

The system is built on a five‑level Agent hierarchy powered by Amazon Bedrock AgentCore (compatible with CrewAI, LangGraph, LlamaIndex, Strands Agents) and AgentCore Memory:

Level 1 – Decision Insight Agent: Combines AgentCore Memory with UserPreferenceMemory to store user preferences, analyze request complexity, and prioritize urgent decisions.

Level 2 – Intelligent Matching Agent: Dynamically adjusts matching weights based on historical shopping preferences and browsing records to surface highly relevant results.

Level 3 – Semantic Compression Agent: Uses advanced semantic encoding to retain 98% of core product information while tripling data‑processing speed and cutting overall processing cost by 80%.

Level 4 – Data Fusion Agent: Applies a proprietary multi‑source data‑cleaning algorithm, achieving an 87% noise‑filtering rate and eliminating cross‑platform information silos.

Level 5 – Personalized Recommendation Agent: Integrates AgentCore Memory to deliver human‑like, emotion‑aware recommendation ordering that aligns with individual shopping habits.

Additional capabilities include:

Intelligent Decomposition: Complex user requests (e.g., “recommend a laptop for a college student under $5,000 that can run design software”) are split into parallel sub‑tasks such as scenario analysis, budget filtering, and software‑requirement matching.

Parallel Routing: Concurrent agents handle different sub‑tasks, reducing overall response time by 60%.

Memory Fusion: Strands Agents retrieve historical user data from AgentCore Memory to synthesize personalized answers that fully reflect unique preferences.

Performance Evaluation

Based on the collaborative architecture of AgentCore Memory and Strands Agents, Monus AI demonstrates a leap in AI search performance, as shown in the accompanying tables (images omitted). The system significantly reduces token consumption for user‑profile‑based searches and improves result accuracy, delivering a competitive edge in both technology and cost efficiency.

Conclusion

The technical partnership between Altre Technology and Amazon Web Services combines a task‑orchestration framework with memory services to answer the core question of how AI can truly understand and serve users in e‑commerce. By leveraging Strands Agents as the skeleton and AgentCore Memory as the brain, along with large‑model collaboration and a semantic consensus engine, the solution establishes a replicable, scalable paradigm that sets a new benchmark for e‑commerce AI search.