Launcheck

LaunchCheck catches conflicting launch decisions before they become real problems.
Engineering wants to delay, Marketing wants to launch, and Support expects a pilot.
GPT-4.1 mini turns company context into role-specific decision scenarios.
Employees independently choose how their team would respond.
Rehearsals are grounded in the conversations teams are already having (e.g. Slack, Gmail)
Connect authorized workplace sources to give AI the context behind decisions.
The CEO chooses an important decision and what the team should stress-test.
Different departments participate in the same company decision.

Inspiration

Big mistakes often come from departments interpreting the same decision differently, not from making the wrong decision.

Before a launch, Engineering knows about an unfixed bug, Marketing has scheduled the press release, and Support is planning a pilot. Each team thinks it's aligned, but together they fail.

We built LaunchCheck to catch this before it happens, inspired by ADP's challenge of turning workplace communication into useful insight.

What it does

LaunchCheck helps companies answer one important question:

Does everyone actually understand the plan, and would they make compatible decisions when something goes wrong?

First, the leader of the company (e.g., CEO) creates a company workspace and connects authorized information sources, such as Slack and Gmail.

Next, the CEO creates a project describing an upcoming business decision, such as launching a new product.

Using GPT-4.1 mini, LaunchCheck analyzes relevant company context and generates department-specific rehearsal questions and scenarios. These questions can explore real concerns or hypothetical situations that could happen in the future.

The CEO can review, edit, and publish the AI-generated questions.

Employees then log into their own accounts and independently answer questions based on their department's responsibilities. They select an action, such as holding the launch, running a limited pilot, or proceeding with a full launch, and explain their reasoning.

Finally, LaunchCheck generates a Readiness Report that compares the departments' selected decisions and highlights disagreements.

For example, in our fictional demo company, Mapleworks:

Engineering recommends delaying the launch because of an unresolved payment bug.
Marketing wants to proceed with the full announcement.
Customer Support expects a limited pilot.

LaunchCheck flags these conflicting plans so leadership can clarify the decision before it affects customers.

How we built it

We built LaunchCheck using Python, FastAPI, SQLite, and JavaScript, with the Slack API and Gmail API connected through OAuth 2.0.

The LLM (OpenAI API with GPT-4.1 mini) writes the scenarios, and backend logic compares the structured answers to find conflicts. We added validation and automatic retries to handle inconsistent AI output.

We also developed a question builder that allows CEOs to review, edit, save, and publish AI-generated questions before employees see them.

Employees access their own dashboards and submit their decisions independently. The backend then compares their multiple-choice actions across departments to identify conflicting launch plans.

Challenges

One of our biggest challenges was getting AI to generate meaningful, realistic scenarios instead of generic questions.

At first, GPT-4.1 mini mostly repeated information it found in Slack. For example, if Engineering mentioned a payment bug, the AI would simply ask whether the team was ready to launch despite that bug. While these questions were useful, they weren't testing how employees would react to unexpected situations.

We experimented with prompt engineering to help the model generate more realistic hypothetical events, such as an accidental launch announcement or pressure from leadership to release a product before it was ready.

We also had to make sure the AI-generated questions made sense for each department and offered clear, consistent answer choices.

Another major challenge was integrating Slack and Gmail securely. Setting up OAuth authentication, handling permissions, and retrieving relevant conversations required us to understand how different APIs work.

We also had to limit how much information the application retrieved so our prototype could run efficiently during the hackathon.

Accomplishments

We built real Slack and Gmail integrations, personalized employee dashboards, a customizable question builder, and a Readiness Report.

Our fictional Mapleworks Checkout demo shows Engineering, Marketing, and Customer Support preparing for incompatible launches, which leaders can catch in advance.

What we learned

We learned that AI has the capability to go beyond summarizing and instead generate realistic scenarios that reveal misunderstandings people wouldn't say out loud.

Pairing an LLM with deterministic logic makes the results consistent. Finding disagreements is only the first step, and the real goal is a shared plan.

We also learned the importance of keeping humans involved. AI-generated questions should be reviewed before publication, and agreement between departments doesn't automatically mean their decision is correct.

What's next

Our next steps include:

Smarter hypothetical simulations: Generate more realistic future crises, unexpected complications, and cross-department scenarios based on company information.
Custom AI instructions: Let CEOs describe exactly what risks or situations they want the AI to stress-test.
AI-powered reasoning analysis: Evaluate whether employees' written explanations support their selected answers and identify contradictions that multiple-choice comparisons might miss.