Security and Compliance
Sanwa Comtec needed to add AI-based harassment detection to on-premises chat deployments for customers that cannot send sensitive conversations to external networks, while still achieving enough accuracy to make alerts useful and compliant.
Reported impact
Sanwa Comtec created a secure, on-premises moderation capability that helps customers detect problematic conversations more accurately while preserving data sovereignty and supporting future production rollout.
Sanwa Comtec built an on-premises harassment-detection capability for Rocket.Chat customers by combining IBM watsonx.ai with IBM Watson natural language processing classification functions. Working with legal experts, the team assembled training examples for sexual and power-based harassment, evaluated multiple models in a proof of concept and selected the most effective approach for a custom classifier. The resulting model runs inside a closed environment so organizations with strict network restrictions can analyze chat content without exposing it to public cloud services. By focusing on lightweight language models that can be packaged on premises, Sanwa Comtec and IBM turned watsonx.ai from a generic model-development environment into a deployable moderation component that fits regulated enterprise communication stacks. The team continues to refine training data and application logic as they move the solution toward production.
Chat messages from on-premises Rocket.Chat deployments are processed by a custom classifier built in watsonx.ai with IBM Watson NLP, which scores conversations for likely harassment and generates alerts without sending sensitive content to external networks.
"> 95％ accuracy of AI-based harassment judgments in closed networks"
"deploying lightweight language models within on-prem packages without relying on cloud-based LLMs"
"issue an alert in a secure on-prem environment that cannot connect to external networks"
"used to train a custom language model within IBM watsonx.ai"
"using the classification capabilities of the IBM Watson natural language processing (NLP) function"
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
IBM Sample Sep 11, 2025 Lenovo PCCW Solutions (LPS) LPS needed a more unified way to monitor, secure, and scale hybrid environments for enterprise clients because its existing monitoring tools only offered siloed snapshots and could not support proactive incident… 24×7 service-level commitments IBM Instana · IBM LinuxONE · IBM Turbonomic IBM Sample Aug 1, 2024 CarbonHelix CarbonHelix needed to deliver highly customized managed security services across diverse customer environments while improving threat-detection speed and accuracy beyond what human analysts and manual workflows could… 24x7 security operations center (SOC) services IBM Guardium · IBM QRadar SOAR · IBM QRadar Suite IBM Sample Jun 7, 2024 boxxe boxxe needed to detect and respond to faster-moving cyber threats across diverse customer environments without overloading its security analysts or relying on fragmented dashboards and manual reporting. 20 hours of manual monthly reporting eliminated IBM Security QRadar SIEM · IBM Security QRadar SOAR 