NYO Automates LLM Gateway FinOps with an Agent on Amazon Bedrock AgentCore

NYO checked its LLM gateway margins by hand with SQL. Swarm and Nyolab built nyo-finops, an agent on Amazon Bedrock AgentCore that writes a daily margin brief and proposes price, limit and credit changes for a person to approve.

FinOps
US$161.77

Unbilled usage found across ~16,700 requests and re-billed

~$730/mo

Estimated losses avoided by repricing a route billed at $0

6 min

From scheduled trigger to the first unattended brief in S3

NYO
Nyolab logo
Case study by
Nyolab

The team behind NYO, a governed AI gateway that puts frontier open models behind one API with per-key guardrails and usage accounting.

Visit hive

About the client

NYO is an LLM gateway that resells model calls for prepaid credits and prices each route per million tokens. It buys tokens from several upstream model providers, including Amazon Bedrock, whose prices change often.

The challenge

Nothing joined upstream cost to what NYO billed, so its owner checked margins by hand with SQL. A route was priced at $0 against a cost of $5.25/$26.25 per million tokens, unbilled usage went unnoticed, and 7 routes had no recorded cost.

The first version of the FinOps agent also ran on a service outside Amazon Web Services (AWS).

The solution

Swarm engaged Nyolab as the specialist Hive team to build nyo-finops and run it in production on AWS. Swarm and Nyolab built the agent with the Strands Agents SDK and deployed it on Amazon Bedrock AgentCore Runtime in eu-west-1. Its model is Kimi K3 on Amazon Bedrock, called through NYO’s route us.kimi-k3.

Every day, Amazon EventBridge Scheduler triggers an AWS Lambda function that starts an agent session. The agent reads NYO’s admin tools over the Model Context Protocol and AWS Cost Explorer data through read-only tools, then writes its brief to Amazon S3.

Amazon Bedrock AgentCore Identity holds its credentials, so the agent holds no AWS keys. Credit and limit changes wait for a person’s approval, and no tool can change prices. Amazon SQS, Amazon CloudWatch alarms and Amazon SNS report failures, and the stack is defined in AWS CDK.

Impact

The agent went live on Amazon Bedrock AgentCore on 27 October 2026. Its first unattended scheduled brief reached Amazon S3 on 28 October, within six minutes of the trigger. A credit grant ran only after the owner approved it.

Earlier in the engagement, the margin engine that Swarm and Nyolab built found US$161.77 of unbilled usage across about 16,700 requests and 8 accounts. Nyolab re-billed it as recovered revenue. The engine also flagged a route priced at $0 against its cost, which was repriced to $6.20/$30.90, avoiding an estimated $730 a month in losses. The agent now uses that engine as a tool.

Nyolab expects more recovered revenue as the agent catches new billing and pricing issues. It also expects more Amazon Bedrock usage as NYO’s router shifts traffic to Amazon Bedrock when providers outside AWS rate-limit requests.

More case studies

Put AI to work across your organization

Book a demo