Sales ops · Production

LinkedIn builds collaborative Jupyter Notebook playgrounds for prompt engineering to develop AI-powered features like AccountIQ

The problem

LinkedIn needed a tight feedback loop bridging engineers doing rapid prototyping and non-technical domain experts, while managing complex LLM configurations, prompt templates with dynamic placeholders, and realistic test data to build production-ready AI features.

Workflow diagram · grounded in source
1
Engineer sets up notebook playground
trigger
“We have added support to launch Jupyter Notebooks from our python code repository. This allows users to clone the repository, build the code and launch the notebooks in browser or IDE with a single command.”
2
Test data collected from data lake
integration
“At LinkedIn, we can query data from our data lake using Trino. Creating good quality test data sets is not a one time task – the data needs to be sampled on a regular basis to get a fresh dataset.”
3
LLM chain orchestration
ai_action
“It internally uses Langchain to orchestrate all the operations, such as fetching and transforming data, preparing the prompts and then calling LLMs. All the prompts to the LLMs are managed using jinja templates.”
4
Real-time Bing Search fetch
integration
“In AccountIQ, we use Bing Search APIs to perform web search to know more about specific aspects of the company.”
5
Sales team prompt iteration
human_review
“This setup has allowed us to give LinkedIn Sales team members access to the prompt engineering playground. Live prompt engineering sessions with end users, observing how they tweak prompts and interact with the system, have provided valu…”
6
Changes committed and deployed
output
“Any changes made to the prompt files can be committed to the internal GitHub repository, versioned, and deployed to production.”
7
Usage data continuous feedback
feedback_loop
“As users interact with our features, we sample the real usage data using tracking. This data helps us better understand the actual usage patterns and also allows us to update the test data sets. This continuous feedback loop helps us con…”
Reported outcome

The Jupyter Notebook playground enabled cross-functional prompt iteration within days and produced AccountIQ, which reduces company research time from two hours to five minutes, achieving significant product-market fit.

Reported metrics
Company research timereducing what used to take two hours to just five minutes
Cross-functional prompt collaboration onboarding timewithin days
AccountIQ product-market fitsignificant product-market fit
Reported stack
Jupyter NotebooksLangchainjinja templatesBing Search APIsTrinogrpcVS CodeIntelliJGitHubOpenAI modelsLlama
Source
https://www.linkedin.com/blog/engineering/product-design/building-collaborative-prompt-engineering-playgrounds-using-jupyter-notebook
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The Jupyter Notebook playground enabled cross-functional prompt iteration within days and produced AccountIQ, which reduces company research time from two hours to five minutes, achieving significant product-market fit.

What tools did this team use?

Jupyter Notebooks, Langchain, jinja templates, Bing Search APIs, Trino, grpc, VS Code, IntelliJ, GitHub, OpenAI models.

What results were reported?

Company research time: reducing what used to take two hours to just five minutes; Cross-functional prompt collaboration onboarding time: within days; AccountIQ product-market fit: significant product-market fit (source-reported, not independently verified).

How is this sales ops AI workflow structured?

Engineer sets up notebook playground → Test data collected from data lake → LLM chain orchestration → Real-time Bing Search fetch → Sales team prompt iteration → Changes committed and deployed → Usage data continuous feedback.

Related sales ops cases