Case study · Commercient Sync · Jan 2024 — Mar 2026
AI agents on top of an ERP-to-CRM integration platform. I built the agents and the plumbing that feeds them.
Commercient Sync connects ERP, CRM and eCommerce systems through a cloud integration platform. Over two years as AI Engineer I turned manual, repetitive work across support, sales and data into agents and self-serve AI tooling that run inside the existing product.
Commercient's value is the data it syncs between ERP, CRM and eCommerce systems. But the work around that data was still manual: support agents looked up orders and invoices by hand, sales reps qualified leads themselves, and internal data questions needed someone to write SQL.
The website chat ran on a third-party bot (Drift) costing about $12k a month, and it could not see any of the business data the platform already had.
A Support Agent that resolves tickets faster by pulling order, invoice and account context from the connected systems.
A Sales Agent, built with LangGraph, that qualifies visitors and generates leads. It replaced the Drift bot.
A SQL View Generator that turns questions into views over synced data.
A bot-builder platform where customers upload their own data from sources like Google Drive, YouTube links and sitemaps, and get a working assistant on top of it.
The scraping tool and ingestion pipelines that feed all of the above.
I loaded and served the LLMs on the company's GPU servers with vLLM instead of relying on per-token APIs.
I designed the agent graphs, the retrieval layer and the fine-tuning runs, and integrated everything into the existing .NET product.
I built the scraping and ingestion tooling, and mentored teammates as the AI surface grew.
.NET, because the product already lived there and the agents had to ship inside it, not beside it.
Python, vLLM, LangGraph, RAG and fine-tuning for the model and agent layer. Each was chosen for the job: vLLM for throughput on our own GPUs, LangGraph for explicit, debuggable agent state.
The Sales Agent replaced a $12k-per-month third-party bot with one that actually knows the business.
Integrating AI into an existing product is a systems problem first and a model problem second.
Mentoring changed how I build: if a teammate cannot debug the agent, it is not done.
Your version of this
Tell me what your team does by hand. I'll tell you what a system for it would look like.
The AI backend that reads a client's tax documents, builds the workpapers and first-reviews the return, on a platform used by 119 CPA firms.