Tellagence Sets New Standard for Contextual Intelligence With First Research-Validated AI Framework

Research demonstrates AI framework improves LLM accuracy and consistency; achieves 96% alignment with human analysis, and up to 30% improvement in consistency across over 433,000 real-world data points

PORTLAND, Ore., May 19, 2026 /PRNewswire/ -- Tellagence, a pioneer in Contextual Intelligence, today announced two research studies validating its proprietary AI framework designed to improve the accuracy, consistency, and reliability of enterprise AI analysis. Published on arXiv and tested across more than 433,000 real-world data points from retail, food service, and publishing, the framework is the first AI preprocessing methodology of its kind to be validated at enterprise scale.

The Tellagence framework serves as a preprocessing layer, transforming unstructured textual data into clean, organized categories that Large Language Models (LLMs) can process more effectively. Rather than feeding LLMs raw text from social media, reviews, and surveys, the Tellagence framework organizes information into a contextual map that helps LLMs focus on the data that matters. The result is reduced variability, improved consistency, and the right context to give a human-level understanding of the conversations shaping business decisions.

"The most dangerous AI output is not one that is obviously wrong – it's the one that is confidently inaccurate," said Matt Hixson, co-founder and CEO of Tellagence. "When AI produces different results each time the same analysis is run, organizations can't tell whether a shift in their data reflects a real change in customer opinion or just model variability. A major strategic decision could be made based on a trend that isn't real. We built this framework to reduce that uncertainty to levels organizations can build strategy on – and the research proves it does."

What the Research Found

Study I: Consistency Analysis of Sentiment Predictions using Syntactic & Semantic Context Assessment Summarization (SSAS)

This study measured whether the same textual data, analyzed repeatedly, produces consistent results when processed by an LLM. It found that the SSAS framework significantly reduced output variability by removing noise and organizing data before it reached the model.

Key findings included:

• Standard LLM analysis produced variability rates between 22–28%, meaning identical inputs generated different outputs across repeated runs
• Applying the SSAS model reduced LLM variability to 1–2.5%
• SSAS noise removal accounted for 20.8–25.6% of the overall improvement
• Up to 30% improvement in the reliability of AI outputs, achieved by removing noise from the data before it reaches the model
• Improvements remained above 20% across all six testing scenarios, including low-volume edge cases

Study II: Leveraging Weighted Syntactic and Semantic Context Assessment Summary (wSSAS) Towards Text Categorization Using LLMs

This study measured whether preprocessing textual data through the wSSAS framework produces more accurate categorization than a direct-LLM approach. It found that wSSAS consistently produced more precise, strategically distinct categories – compared to the fragmented, overlapping results of a direct-LLM approach with no preprocessing."

 Key findings included:

• wSSAS generated more precise and strategically distinct categorization structures
• In comparison, Standard LLM categorization produced fragmented and overlapping categories
• Applying wSSAS improved cluster integrity by 163% within the Google Business Reviews dataset
• Across all datasets tested, wSSAS delivered equal or better categorization performance in 80–92% of evaluated segments

"The problem with most AI analysis isn't the model – it's what the model receives," said Nitin Mayande, co-founder and Chief Science Officer of Tellagence. "We developed this framework to solve that, giving the AI a structured map to navigate while directing its attention to what actually matters. These findings demonstrate that cleaner input produces more consistent output – which gives organizations intelligence they can actually act on."

Additional Benchmarking

Separate internal benchmarking conducted by Tellagence showed 96% alignment with human analysis while processing 100% of evaluated datasets without sampling bias. The framework completed the analysis in approximately 20 to 120 minutes, compared to the weeks required by conventional approaches.

"What makes this framework significant is not just the results, but the ability to reproduce them consistently at scale," said Dr. Nitin Joglekar, study co-author and advisor at Tellagence. "We validated the framework across multiple independent datasets representing different industries, and the findings remained consistent."

How the Studies Were Conducted

Both studies have been conducted using Gemini 2.0 Flash Lite across three independent, large-scale real-world datasets sourced from the University of California, San Diego: approximately 155,000 Amazon Product Reviews (retail), 121,000 Google Business Reviews (food service), and 157,000 Goodreads Book Reviews (publishing).

The framework has been evaluated across 10 independent runs and six robustness scenarios, spanning optimal data conditions to extremely low-volume edge cases.

The studies are co-authored by Shreeya Verma Kathuria, Nitin Mayande, Sharookh Daruwalla, Nitin Joglekar of Villanova School of Business, and Charles Weber of Portland State University. Both papers are published in Computation and Language (cs. CL) and Artificial Intelligence (cs.AI) on arXiv.

Enterprise Applications

For organizations using AI to analyze customer insight, track brand perception, or identify drivers behind customer behavior, the new framework is used for a number of applications:

• Intelligence and research: survey analysis, brand perception, cultural trend identification, and social media insights
• Strategy and planning: audience analysis, brand differentiation, market positioning, and product or service opportunity identification
• Execution and monitoring: campaign performance, influencer vetting, crisis management, and real-time signal detection

The framework is accessible through Tellagence's AI platform via an interactive dashboard, automated Pulse Reports, and an API layer supporting 140 languages, or through Tellagence's tailored services where experts work alongside teams to surface the insights needed.

About Tellagence

Tellagence is a pioneer in contextual intelligence, partnering with leading agencies and global brands to transform conversations into trusted, actionable intelligence. Founded in 2011 on the belief that language only carries meaning in context, Tellagence has spent more than a decade developing an AI ecosystem that delivers human-level understanding of the conversations shaping business decisions.

For more information, visit https://www.tellagence.ai/

Press Contact:
Annika Jensen-Lamka
[email protected]
Mobile 415-516-3530

SOURCE Tellagence