Achieving 93% Consistency: How Narwal Built a Production-Ready AI Sentiment Assistant for a Global Manufacturer

Background

A leading global manufacturer operating across a diverse portfolio of building and consumer products needed an AI sentiment assistant to make sense of large volumes of customer reviews across retail, distributor, and digital channels throughout North America. Teams needed a faster and more reliable way to analyze this feedback, identify emerging product issues, and surface actionable insights at scale.

To support this initiative, the company partnered with Narwal to design and implement a Snowflake-native AI solution that could improve retrieval accuracy, enable conversational querying of review data, and operate fully within existing enterprise governance boundaries.

Challenge

The organization identified several gaps in their existing approach to customer feedback analysis:

• High Volumes of Unstructured Feedback: Product reviews across channels continued to grow, making manual analysis slow, inconsistent, and difficult to scale.
• Low Retrieval Precision: Existing keyword-based search surfaced technically related but contextually irrelevant reviews, reducing confidence in downstream insights.
• Unvalidated AI Responses: The organization lacked a mechanism to verify whether AI-generated responses were accurate, complete, and grounded in actual customer review data.
• Limited Self-Service Access: Product and business teams relied heavily on engineering support to query and interpret customer feedback datasets.
• Governance and Data Residency Requirements: External AI platforms were not viable due to enterprise governance, compliance, and data residency constraints within the Snowflake ecosystem.

Solution

Narwal designed and implemented a Snowflake-native Product Sentiment Assistant using a hybrid RAG architecture optimized for retrieval quality, governance, and usability:

• Hybrid Retrieval Pipeline: Dense vector retrieval combined with keyword filtering and Cortex Search re-ranking improved the relevance of retrieved reviews. Introducing re-ranking significantly improved contextual precision over embeddings alone.
• Snowflake-Native AI Execution: Cortex Search and Cortex Complete handled retrieval and LLM inference directly within Snowflake, eliminating the need for external AI services or data movement outside the governed environment.
• LLM-as-a-Judge Validation Loop: A secondary LLM evaluated every generated response for correctness and completeness before delivery, directly reducing hallucination risk.
• Streamlit-Based Conversational Interface: A Streamlit application with session management, chat history, and visualization dashboards enabled product teams to explore customer feedback conversationally without engineering intervention.
• Snowflake-Only Architecture: All compute, storage, and AI inference remained within Snowflake, satisfying enterprise data governance and security requirements.

Outcomes

The solution delivered measurable improvements across retrieval quality, response reliability, and business usability:

• 93% Response Consistency: The LLM-as-a-Judge validation layer improved response reliability across repeated interactions and evaluation runs.
• 83% Response Completeness: Responses were more comprehensive and better grounded in review data, reducing ambiguity for product decision-makers.
• Reduced Hallucination Risk: Low-confidence or weakly grounded responses were identified before reaching end users.
• Real-Time Feedback Querying: Product teams moved from manually reviewing large volumes of customer comments to querying thousands of reviews conversationally in real time.
• Production-Ready Delivery: Narwal delivered a complete proof of concept including retrieval pipelines, validation workflows, and a Streamlit-based user interface ready for enterprise deployment.

Conclusion

Narwal delivered a production-ready AI sentiment assistant that gave a global manufacturer’s product teams direct, reliable access to customer intelligence at scale. Built entirely within Snowflake, the solution met strict governance requirements while achieving strong evaluation metrics across retrieval precision, response consistency, and completeness. Product teams gained a self-serve capability to turn large volumes of unstructured feedback into faster, more confident product decisions.