🚀 TechFutures 2025 — End-to-End MLOps on Databricks

Welcome to the TechFutures 2025 hands-on workshop:
"End-to-End MLOps Pipelines on Databricks: From Model Training to Production."

Led by Debu Sinha — Lead Applied AI/ML Engineer at Databricks and author of
Practical Machine Learning on Databricks.

---

🧠 Workshop Overview

In this workshop, you'll learn how to build and scale a complete MLOps lifecycle on Databricks:

• Train and track models using PyTorch + MLflow
• Register and govern models with Unity Catalog
• Deploy and serve models using Model Serving / AI Gateway
• Automate experiments with Optuna and MLflow nested runs
• Extend the same workflow to LLMs via Mosaic AI

---

📚 Notebook Index

#	Notebook	Description
1	01_MLOps_Pipeline_on_Databricks.py	Hands-on notebook showing an end-to-end pipeline — train, track, register and deploy a PyTorch model with MLflow.
2	02_Advanced_MLflow_PyTorch_Tutorial.py	Advanced workflow covering Optuna tuning, Unity Catalog governance, and Spark batch inference.
(Optional) 3	03_LLMOps_with_MosaicAI_and_MLflow.py (coming soon)	Extends the same lifecycle to LLMs with prompt tracking and Mosaic AI Serving.

---

🏗️ MLOps Lifecycle on Databricks

Pipeline Stages

1. Data Preparation (Delta Tables)
2. Model Training (PyTorch + MLflow)
3. Experiment Tracking (MLflow UI)
4. Model Registry (Unity Catalog)
5. Model Serving (REST API / AI Gateway)
6. Monitoring (Lakehouse Monitoring + Alerts)

---

🤖 LLMOps Lifecycle — Unifying Generative AI Workflows

Lifecycle Flow

1. Data (Prompts & Responses)
2. Fine-Tuning (Mosaic AI Training)
3. Tracking (MLflow Metrics + Prompts)
4. Registry (Unity Catalog Models)
5. Serving (Databricks Model Serving)
6. Evaluation (LLM-as-a-Judge, Metrics, Drift)

🧩 MLOps → LLMOps → AgentOps — a unified Databricks AI platform.

🔬 Hands-On LLM Fine-Tuning Demos

To bridge the gap between traditional MLOps (Notebooks 1-2) and the upcoming LLMOps notebook, explore these interactive fine-tuning examples that demonstrate the same lifecycle principles applied to Large Language Models:

📖 Databricks LLM Fine-Tuning Demos

These demos showcase:

• Classification Fine-Tuning — Specialize models for customer support scenarios using the Chat API
• RAG Chatbot Fine-Tuning — Optimize conversational AI with retrieval augmentation and conversation history
• Entity Extraction Fine-Tuning — Train models for specialized extraction tasks (e.g., medical/drug entities)
• MLflow Evaluation — Measure fine-tuning improvements with built-in evaluation frameworks

Why These Demos Matter:

• Apply the same MLflow + Unity Catalog patterns from Notebooks 1-2 to LLM workflows
• See how model tracking, versioning, and serving extend to generative AI
• Practice evaluation techniques specific to LLMs (including LLM-as-a-Judge)
• Bridge traditional MLOps foundations with modern LLMOps practices

These demos complement Notebook 3 (coming soon) and provide immediate, hands-on practice with production LLM workflows on Databricks.

---

⚙️ Setup & Requirements

• Databricks workspace with MLflow & Unity Catalog enabled
• Runtime: Databricks Runtime ML 15.x or above
• Python 3.10 +
• Optional: Mosaic AI Model Serving access

---

🚦 Run the Notebooks

Clone this repo or import directly into Databricks:

git clone https://github.com/debu-sinha/techfutures-2025-mlops-databricks.git

---

📖 Additional Resources

• Practical Machine Learning on Databricks — Debu's comprehensive guide
• Databricks MLOps Workflows — Official documentation
• The Big Book of MLOps — Free comprehensive guide

---

Questions? Reach out during the workshop or connect with Debu on LinkedIn

---

Made for TechFutures 2025 attendees