2026 • Data Engineering Internship

Waldinger LLM Pipeline

Explore Project Visit Waldinger

0 %

Cost Reduction

Idle compute costs eliminated

0 +

Snowflake Objects

Schema objects monitored for drift

Model Parameters

Mistral-7B-Instruct deployed on Azure ML

Inference Modes

Real-time + batch inference supported

Project Overview

Waldinger's data platform ran Snowflake across 3,000+ objects with no automated validation, and Azure ML compute clusters sat idle at peak provisioning between jobs. Two fixable problems. This internship was about fixing them.

The core deliverable was a production LLM inference pipeline on Azure ML using Mistral-7B-Instruct, supporting both real-time and batch modes from a single registered model artifact. Infrastructure was written entirely in Terraform so deployments are reproducible and configuration drift is structurally impossible.

The schema drift detection system runs on every pipeline execution, snapshots Snowflake object definitions, and surfaces deviations before they cascade into downstream failures. Between that and compute right-sizing, idle spend dropped 70%.

My Role

LLM Pipeline Engineering

Built Mistral-7B-Instruct inference on Azure ML supporting real-time and async batch modes from one registered model artifact, versioned through Azure ML Model Registry.

Infrastructure as Code

Provisioned Azure Container Apps, ML Workspace, and ACR entirely in Terraform. Every resource is declarative, reproducible, and immune to manual drift.

Snowflake Drift Detection

Built a system that snapshots definitions for 3,000+ Snowflake objects on each pipeline run and alerts the data platform team the moment a deviation is detected.

CI/CD Automation

Wired model training, registration, and deployment into GitHub Actions workflows so the entire path from code commit to live inference runs without human intervention.

Technical Stack

AI / ML

Mistral-7B-Instruct Core LLM model

Azure ML Training & hosting

Azure ML Model Registry Versioned artifacts

Python Pipeline & scripts

Infrastructure

Terraform IaC provisioning

Azure Container Apps Serverless containers

Azure Container Registry Image management

GitHub Actions CI/CD workflows

Data Platform

Snowflake Cloud data warehouse

Drift Detection 3,000+ object monitoring

SQL Schema introspection

Data Pipelines ETL orchestration

Challenges & Solutions

Compute Cost Overrun

Azure ML clusters were provisioned at peak capacity and left running between jobs. No scale-down policy, no idle detection, just continuous spend.

Fix. Auto-scaling policies via Terraform with min/max instance bounds, combined with Azure Container Apps scale-to-zero for the serving endpoint. Idle compute costs dropped 70%.

Snowflake Schema Drift at Scale

Three thousand Snowflake objects, zero automated validation. Ad-hoc schema changes went undetected until something downstream broke.

Fix. A drift detection layer that snapshots object definitions on each pipeline run and diffs against expected baselines, surfacing discrepancies before they can propagate.

Conflicting Inference Latency Requirements

Teams had opposite needs: some wanted sub-second real-time responses, others needed cost-efficient overnight batch processing. One model, two incompatible serving patterns.

Fix. Two deployment targets from one registered model artifact. An Azure Container Apps endpoint for real-time calls, a batch cluster for async jobs. Same model, different serving behavior, managed from one Terraform configuration.

Key Learnings

Ninety percent of production ML impact isn't the model. It's what runs around it: versioning, infrastructure, cost controls, observability. Ship those right and the model mostly takes care of itself.

Terraform isn't just provisioning. When infrastructure is code, a misconfigured cluster becomes a code review failure rather than a runaway bill.

Schema drift detection is only useful if it runs before the pipeline, not after. Detection timing is the entire design problem.

Scale-to-zero isn't just a cost strategy. It forces you to design stateless serving, which improves the architecture beyond the billing savings.

Model registry versioning changes how teams collaborate. When every artifact has a version and a deployment history, rollback stops being a crisis.

Explore More

Waldinger Corporation

Visit the company website

→

All Projects

Next Project Dwolla Dashboard