AI Native ML Tool

Hiring your first Machine Learning Engineer is a catch-22 for midsize companies.

Product teams want to add intelligent features—like churn prediction or dynamic pricing. But Leadership won't approve the $150k+ headcount until they see proof that the feature is technically feasible and holds real business value.

Without the MLE, you can't build the Proof of Concept (PoC). Without the PoC, you can't hire the MLE.

This is the story of how the Product Manager (Alex) and the Engineering Lead (Jordan) at a midsize B2B SaaS company used Gaslamp to break the deadlock on a Friday afternoon.

The Mandate: Churn Prediction

The company had a retention problem. The CEO wanted to know if they could predict which users were going to churn before they canceled.

Alex and Jordan knew the data existed—they had gigabytes of raw product usage logs. But digging through unstructured logs to build a predictive model wasn't a weekend side project.

Or so they thought. They fired up a terminal and invoked Gaslamp.

/gaslamp We want to build a churn prediction model to save at-risk accounts.

Phase 1: Clarifying the Business Problem (The Interview)

Gaslamp didn't immediately ask for a CSV file. It started an interview.

[Gaslamp] Churn can be tricky to predict if the signal is too late. How are we defining the "Churn Event"?

Is it the moment they click "Cancel Subscription"? (Lagging indicator - hard to save them).

Or is it "Zero logins for 30 consecutive days"? (Leading indicator - actionable).

Alex and Jordan had originally planned to use cancellation dates. Gaslamp's pushback forced them to pivot to zero logins for 30 days—a much more actionable metric for the Customer Success team.

Artifact Created: project_brief.md

Phase 2 & 3: The $10 Model (Data & Architecture)

Jordan exported a sanitized 10GB sample of their raw event logs.

[Jordan] We have the raw logs. Do we need an AWS GPU instance to process this and train a neural network?

Gaslamp analyzed the schema.

[Gaslamp] No GPUs required. We don't need Deep Learning for tabular behavioral data.

Data Strategy: Let's aggregate the raw events into weekly counts (e.g., logins_last_7d, features_used_last_14d).

Architecture: We will use a Random Forest Classifier. It's fast, interpretable, and trains on a standard CPU in under a minute.

Action: Generating preprocess.py and train.py...

It took Gaslamp about $4.50 worth of LLM API tokens (its "reasoning" cost) to write the architecture and the data pipeline. The model trained locally on Jordan's laptop.

Phase 4: Proving Feasibility (The Eval)

Ten minutes later, the training loop finished.

[Gaslamp] Evaluation Report:

Precision: 71% (When we flag an account as at-risk, we are right 7 out of 10 times).

Recall: 65% (We catch nearly two-thirds of all churning users).

Verdict: PASS. This is a strong starting signal for an Alpha release.

It wasn't a perfect, state-of-the-art model. But it was a working, end-to-end predictive pipeline.

Artifact Created: eval_report.md

The Outcome: Budget Approved

On Monday morning, Alex and Jordan walked into the leadership meeting. They didn't pitch an abstract idea; they presented Gaslamp's eval_report.md.

They proved that the raw usage data contained a highly predictive signal, and they showed a working alpha model—all built for less than $10 in API costs and a few hours of Friday afternoon time.

The CEO approved the budget for the MLE requisition that afternoon.

Gaslamp isn't here to replace the MLE. It's the catalyst that gets the MLE hired by proving the initial business case for pennies.

> Note: While this story is based on a real customer use case, the names of the company, brand, and individuals have been changed to protect privacy.