How CBMG Slashed Server Costs by 50% and Taught AI to Deliver Personalized Feedback
STX Next, a company specializing in software modernization and AI, partnered with Cohen Brown Management Group (CBMG) to upgrade their legacy B2B corporate training platform.
Our engineering team built a scalable AI speech evaluation microservice and optimized the cloud infrastructure, cutting server costs by 50%. This project outlines the technical steps for integrating gen AI, addressing legacy software migration and optimizing cloud costs.
Moving from manual self-assessments to automated AI audio scoring
Two years ago, trainers at CBMG were fighting against their own software. Building a single training module was an exercise in frustration, a tedious journey through a maze of nested menus and form fields.
The platform's core value proposition was fundamentally broken. The system was designed to evaluate whether a trainee had mastered a complex sales conversation, yet the grading was left entirely to the trainees themselves. They simply ticked boxes to confirm they had hit the right talking points.
CBMG needed an automated performance evaluation system they could trust and that would provide real support in upskilling the team. And they needed an AI-experienced partner who could carry out this project.
STX Next’s team fundamentally re-engineered CBMG's core platform. We cut infrastructure costs by over 50% and introduced a new advanced AI system. This system analyzes human speech, evaluates tone, timing, and confidence, and gives personalized feedback. It does this much more reliably than self-reporting ever could.
A Four-Decade Veteran of Corporate Training
For over forty years, Cohen Brown Management Group (CBMG) has built training programs for some of the largest corporate entities globally.
Their expertise lies in what they call techno-behavioral solutions – software platforms built to train employees in the critical, client-facing behavioral skills that drive sales and customer service. Their flagship products, Cbway and ProPractice, deliver this behavioral training at massive scale across global enterprise environments.
High Infrastructure Costs and Unreliable Performance Metrics
When CBMG partnered with STX Next, their engineering and product roadmap faced three immediate operational challenges:
Unpredictable Reliability: The legacy module creation platform suffered from stability issues that disrupted daily workflows.
Data Blind Spots: There was no reliable way to see how trainees were interacting with the system, leaving their analytics initiatives without clean data.
High Operational Drag: Running the legacy infrastructure was too expensive. It burned through a budget that could have been used for R&D instead of covering server bills.
Beneath these operational issues was a bigger problem with their main product, ProPractice. The system depended on trainees to report their own performance, so the data was often unreliable and could be easily manipulated. Accidental clicks and naturally inflated self-scores quietly eroded the integrity of CBMG’s data. This damaged internal analytics and jeopardized the credibility of the ROI reports CBMG sent to its corporate clients.
A Rebuilt Platform and a Suite of AI Scoring Systems
We approached the problem on two distinct fronts: fixing the immediate developer and user experience, and completely reimagining how human performance is measured.
A Faster, Intuitive Module Builder
Working closely with CBMG’s Product Design team, we replaced the old administration interface with a streamlined frontend built on Angular and DaisyUI. By eliminating unnecessary cognitive load and simplifying the form workflows, we cut the time required to build a single training module by over 20 minutes.
Automated Speech Evaluation for ProPractice
We replaced the unreliable checkbox system with an intelligent, audio-first evaluation engine powered by Gemini. When a trainee records their response to a sales scenario, the AI listens, transcribes the speech, and automatically verifies whether they covered the necessary behavioral markers.
Crucially, the system goes beyond simple keyword matching to perform a qualitative style analysis, evaluating:
- Rhythm and timing
- Volume and tone
- Emotional authenticity and confidence
The backend for this core engine was built using Python and Django, paired with an Angular frontend.
A Dedicated Feedback Microservice
To keep the architecture clean and loosely coupled, we isolated the sales training analysis into a standalone microservice. This service accepts raw audio, transcribes it, checks for required content, and generates contextual coaching tips to help the trainee improve.
By building this with FastAPI, Pydantic, Langfuse, Docker, and Poetry, and deploying it on GCP Cloud Run, we ensured the service can scale horizontally to handle massive spikes in user traffic without driving up idle infrastructure costs.
Validating Behavioral Analysis (PoC)
To explore the boundaries of automated grading, we built a proof of concept (PoC) that evaluates entire training session transcripts against complex compliance and scenario requirements. Using large language models from Azure OpenAI, this prototype scores responses on style, correctness, and exact sequence alignment. We delivered a comprehensive technical report alongside this PoC, outlining the dataset analysis and the resource models required for enterprise scaling.
An AI Assistant for Content Creators
For CBMG’s call center training platform, we built an interactive script-generation assistant using LangGraph, Gemini, Streamlit, and FastAPI. This tool guides trainers through a conversational workflow to build training scripts from scratch.
To ensure the output remains safe and highly relevant, we implemented programmatic input and output guardrails:
- User input
- Input guardrails: rejects out-of-scope prompts
- LangGraph + Gemini engine: asks clarifying questions if context is missing
- Output guardrails: programmatically checks script requirements
- Final actionable script: or auto-revision if requirements fail
Disciplined Product Management
Building advanced AI features matters very little if the project slips past its deadline. We introduced a structured way of working to the CBMG team, breaking down the roadmap into independent, INVEST-compliant user stories. This allowed us to parallelize development from day one. Backed by rigorous schedule feasibility assessments, regular retrospectives, and a comprehensive regression testing plan, the project was delivered on time and on budget, covering 95% of the initial scope.
Tech Stack
Cutting Server Costs by 50% and Module Creation Time by 20 Minutes
Transitioning to a modern, AI-powered architecture delivered immediate results across the organization:
Infrastructure Savings
Cut ongoing server bills by more than half, freeing up capital for continuous platform improvements.
Operational Efficiency
Saved administrators 20+ minutes of manual work per training module through a cleaner UI workflow.
Data Integrity
Eliminated the gamed self-scoring system, replacing it with objective, automated AI metrics that corporate clients can trust.
Scalable Coaching
Shifted the platform from a passive grading tool to an active learning partner, delivering real-time, highly personalized feedback to trainees.
No More Bottlenecks
Removed the need for human-in-the-loop grading, allowing trainees to complete modules asynchronously and at their own pace.
Predictable Delivery
Completed a major platform and AI overhaul on time, on budget, and within 90% of the target scope.
Summary
CBMG came to STX Next with a legacy platform weighed down by high infrastructure costs, reliability issues, and an operational bottleneck built on an honor system.
Over a two-year partnership, a small engineering team stabilized their foundational software, halved their cloud spend, and introduced a sophisticated suite of AI tools. By shifting from subjective checkboxes to objective, Gemini-powered speech analysis, CBMG now possesses a modern training platform capable of delivering personalized professional development at the global scale their brand requires.
Looking to integrate generativeAI into your product?
Schedule a call with our experts to discover the benefits of partnering with STX Next.
Our customers love to work with us
