AI Engineering Platform

Problem

At Latitude, we struggled to create reliable prompts for our previous product: a data chatbot. We kept iterating on prompts without a clear way to track changes or evaluate results, even though the data required 100% accuracy. Interviews with other product teams revealed similar challenges when building AI features.

• LLMs are not deterministic — They can't guarantee accurate results for every request and may generate incorrect information.
• Small changes affect results — Editing a prompt can improve some use cases while making others worse.
• Collaboration is hard — When prompts live in code, domain experts and other non-technical team members can't edit them directly.
• Testing is difficult — Teams struggle to test AI workflows across use cases and large volumes of real data before deployment.
• Limited observability — Teams lack visibility into prompt performance in production and ways to test prompts with real data before deployment.
• Fragmented version control — Each company tracks prompt versions differently, using tools such as Excel, Notion, and Google Docs.

Opportunity

Our research revealed common challenges across companies working with LLMs, a rapidly growing technology. The global LLM market is projected to reach $259 billion by 2030.

The market was still young, and we saw a gap in tools for managing prompts. We pivoted to focus Latitude on prompt engineering and help teams work more confidently with AI.

Solution

We organized the solution around three stages: build, evaluate, and refine.

First, we made prompts easy to build and test. Then, we gave users tools to evaluate performance. Finally, we helped them use those results to refine prompts and improve accuracy.

Projects and Version Control

We started with collaboration and version control, organizing Latitude around projects and a versioning system similar to Git.

Each project can contain multiple prompts: files with instructions for an LLM. Versions preserve the history of changes to those prompts.

Developers and other team members can edit prompts together in a draft before deploying them. Once deployed, the prompts are available to their apps through our API or SDK.

Editor

The second step was creating an editor to build and edit prompts. Users have access to a simple tag syntax to define system, user, and assistant messages.

We supported more complex workflows with prompt references, parameters, loops, and conditions. Users can also chain prompts to pass results from one prompt to another.

A live preview lets users test prompts and see results as they edit.

Evaluations

With building and collaboration covered, we focused on helping users understand prompt performance.

We introduced evaluations: prompts that analyze another prompt's output and return a score. Users can assess quality before and after deployment with test data and production logs.

Refiner and Copilot

The final step was turning evaluation results into improvements. Once users identified a poorly performing prompt, they needed a way to refine it.

We built a refiner that runs a series of prompts to improve the original using low-scoring evaluation results. Users can then test the revised version without starting from scratch.

We also added Copilot, an LLM-based assistant that understands the platform's syntax and capabilities and helps users write and improve prompts.

Success

We validated the solution in three steps: interviews to understand users' problems and context, prototype sessions to gather feedback and iterate, and a beta to test the product with real users and track usage.

In the first month, we had more than 1,000 registered users and a retention rate of 10% to 20%. Our next priority was to improve retention through further iteration.