Collinear streamlines model customization using Amazon SageMaker AI

As agentic AI becomes more common in enterprise environments, companies use RL to improve agent behavior and capabilities. Collinear addresses this trend with simulation lab environments for training, testing, and post-training agents before production deployment. These environments use large language models, such as Llama, Qwen, and DeepSeek, to power simulated coworkers and users—known as NPCs—that agents interact with while completing tasks.

Creating realistic RL environments is complex. Each NPC must represent a different kind of user, such as a helpful marketing colleague or a frustrated IT employee. Collinear’s initial approach to training AI agents involved fragmented frameworks that required hands-on involvement from multiple team members. “The lifecycle involved planning each experiment, setting it up across different screens and interfaces, tracking it, and then finalizing it and getting the results,” says Soumyadeep Bakshi, cofounder and chief product officer at Collinear AI.

Training cycles took several weeks due to extensive setup and coordination. When AWS introduced its latest serverless tuning experience, Collinear saw an opportunity to unify and refine its training stack within a single solution.

AWS solutions architects introduced Collinear to Amazon SageMaker AI, a fully managed service that brings together the most comprehensive set of AI tools. With SageMaker AI, Collinear could accelerate its entire model customization workflow, from data preparation and technique selection to agent training, evaluation, and deployment.

Along with its ability to fine-tune the large language models that power NPCs, Amazon SageMaker AI offered a significant boost in experimentation speed—a cost-saving factor for Collinear. “Time is an invaluable asset for a startup creating simulation labs for high-profile customers,” says Bakshi. “The increased experimentation speed of Amazon SageMaker AI was the biggest value add.”

Another advantage for Collinear is the serverless infrastructure available through Amazon SageMaker AI. This removes the need to manage training infrastructure and supports a unified, complete workflow within a single interface. As a result, Collinear gains improved visibility and oversight across the entire training lifecycle.

Within Collinear’s architecture, customer agents enter the company’s simulation environment and interact with NPCs representing enterprise users. As agents complete tasks, such as responding to messages or updating records, the environment documents detailed interaction data and outcomes. These interaction signals help identify where an agent succeeds or fails. Customers can then use those signals during post-training workflows on AWS to refine the agent’s behavior and improve performance.

Collinear was already one of hundreds of thousands of customers running AI workloads on AWS. “Collaborating with the AWS team to find an AI model customization solution was the best way to consolidate our lifecycle,” says Bakshi. “They’ve consistently helped us drive innovation on our product, and they’re super technical, helping us stay 3–6 months ahead of the leading AI labs.”

By implementing Amazon SageMaker AI, Collinear reduced its experimentation cycles from 2–3 weeks to 2–5 days, representing up to 90 percent time savings. Its RL environment and NPC simulations now achieve roughly 90 percent real-world fidelity.

Moving forward, Collinear aims to further improve the realism of its RL environment, with a long-term goal of reaching 99 percent fidelity to real-world use cases. Achieving this would help the company to simulate hundreds of users within an organization—including human resources, finance, sales, marketing, and other departments—each represented by NPCs with distinct personalities and behaviors.

It’s an ambitious goal, and one Collinear trusts AWS to help it achieve. “Building RL environments is hard, but our AWS collaboration has helped us better support enterprise customers,” says Bakshi. “Using AWS, we’re able to complete simulations of real-world environments.”