We developed this solution after a B2B client ran into issues deploying production AI agents to speed up their customers’ onboarding.

Their UX: a customer enters their company homepage URL, and the agent autofills key fields (logo URL, case study URLs, etc.) to reduce manual form-filling and drop-offs.

The first version used ChatGPT and Claude in a relatively simple agent workflow. Once they started correcting frequent mistakes in critical fields (especially company logo URL and case study URLs), the system got more complex: • de novo crawling of the homepage URL • cross-checking every LLM response • multiple LLM calls per onboarding

Tokens per API call skyrocketed, pushing costs to about $0.80–$2.00 per call at ~1,000 queries/day, which wasn’t sustainable.

Our solution uses a combination of smaller language models running locally that consume outputs from pre-validated code snippets (for extracting URLs and other company details). A “reasoning” LLM selects which snippets to run at inference time based on the homepage content. This keeps most of the heavy lifting in deterministic code, with the LLM mainly orchestrating.

On a small initial test set of 30 URLs from the client, we returned accurate field values for 27/30. The workflow is noticeably faster end-to-end, and we charge a flat $0.25 per call (including crawling + extraction). The agentic workflow is exposed via a WebSocket interface (JSON in/out). You can see it in action here, along with a free tester API key in the notebook: Google Colab: https://colab.research.google.com/drive/1AkLpL6IQoMDt6-aJwhT...

I’d love feedback, especially on: • obvious failure modes we’re not accounting for • fields you’d expect beyond those in the colab • whether per-call pricing vs. per-token pricing makes sense If you want to try this in your own onboarding stack or have ideas for features, you can reach us at contactus@skymel.com.