Genius Solutions - Genius ERP

Our focus split across two tracks: predictive models and the agentic assistant.

Part One: Predictive Models

To create machine learning models that could make accurate predictions for manufacturers using the rich data (already collected in the ERP), we targeted three high-impact use cases: Vendor Lead Time, Job Delays, and Demand Forecasting.

Use Cases

• Vendor Lead Time - Predicts whether a supplier will deliver late based on historical data. For example, if a battery module supplier regularly misses deadlines, the ERP can flag it early, helping EV manufacturers adjust build schedules or find alternatives.

• Job Delays - Anticipates delays in production jobs caused by a variety of factors, including supply issues, planning gaps, machine downtime, or human-related challenges. The ERP uses historical data to flag risks early, helping teams adjust schedules and keep operations on track.

• Demand Forecasting - Estimates future inventory needs. If next month’s forecast shows a spike in demand for a specific EV model, the ERP calculates what parts (such as charging units or chassis components) need to be restocked and when.

Before this project, Genius ERP offered light automation, like reorder reminders, though most decisions were still manual. Our goal was to embed predictive intelligence into the workflow. We used traditional supervised ML models for lead time and job delay predictions. For demand forecasting, we built a hybrid solution combining Chronos (a time-series forecasting foundation model), classical ML, and statistical methods.Each client has their own database, so we created an automated pipeline using Azure Machine Learning. This securely fetches client data, trains models in the cloud, and deploys them back into the environment, keeping everything up to date without manual work.

Part Two: Agentic Assistant

In the second part of the project, we laid the groundwork for an agentic assistant, named Cortex, that could serve as an always-on, in-platform expert. Cortex uses natural language to query a deep knowledge base sourced from two rich content hubs: Genius’ Online Help Guide (which includes product definitions and terminology) and the Genius Academy, a learning platform full of video tutorials, quizzes, images, and other educational content.

We created Extract, Transform, Load (ETL) pipelines to pull that content, clean it up, and convert it into a latent format the assistant could understand. First, we “cracked open” documents in different formats like HTML, video, and PDF, extracting the full text with semantic meaning. From there, we used embedding models to convert the text into vector representations that can be stored and searched efficiently.

Some of the video content was over two hours long, so we developed workflows to break up long files, extract audio, process it using Whisper for transcription, and prepare it all for ingestion. To make this sustainable, we also introduced a Change Detection Feature. This identifies exactly what’s changed in the knowledge base content since the last pipeline run, so only new or updated material is processed (saving time and compute).

Because failures can happen when running pipelines, like API rate limits or service interruptions, we used a blue-green deployment strategy for the search index. That means we keep the current version live while building a new one in parallel. Once the new version is validated, we switch over instantly with no downtime.

At the application level, each ERP deployment has its own agent with access to a dedicated toolset. The most important of these is the knowledge base itself, which users can query in natural language. We built this using Semantic Kernel, Microsoft’s orchestration framework for integrating large language models into .NET applications.