Milestone 2: MLOps Infrastructure

Key Dates

  • Due: 10/20

Overview

This milestone focuses on moving from proposal to implementation. You will set up reproducible working environments, containerize core components, and establish foundational data operations.

  • LLM teams: Build an initial RAG-ready workflow (data collection, chunking, and vector database integration at a basic functional level).
  • Vision / other teams: Build a robust data ingestion/preprocessing workflow and train or adapt a baseline model.

By the end, each team should have the required components in place and an initial working app skeleton informed by the Milestone 1 mock-up.

Template Repository

[ADD LINK]

Objectives

Build a reproducible, containerized pipeline and establish practical data management and application foundations.

1. Virtual Environment Setup

  • Create virtual machines/environments to support containers.
  • Provide clear setup documentation.

2. End-to-End Containerized Pipeline

  • Containerize each component (e.g., ingestion, preprocessing, RAG steps).
  • Compose them into a single pipeline that runs start-to-finish.
  • The pipeline must be runnable with one command (e.g., docker compose up or make run).
  • Include basic data/version tracking (e.g., dataset snapshot IDs, file versions, or a lightweight versioning tool).

3. Teams Using LLMs

  • Implement an initial RAG pipeline with data collection, chunking, and vector database integration.
  • Show at least one complete retrieval flow from query to retrieved context.

4. Teams Using Vision or Other Modalities

  • Create a robust data ingestion and preprocessing pipeline.
  • Train or adapt a baseline model for your task.
  • Document early experiments and rationale for next technical decisions.

5. Application Mock-up

  • Submit a simple working app skeleton or clickable mock-up.
  • Refine or extend the Milestone 1 design based on feedback and what you learned during implementation.

Deliverables

1. Virtual Environment Setup

  • Screenshot of running instances (cloud or local).

2. End-to-End Containerized Pipeline

  • Dockerfiles + build instructions.
  • pyproject.toml (using uv) for each container.
  • Scripts or docker-compose.yml (when applicable).
  • Documentation explaining the pipeline and exact run instructions.
  • Evidence it works end-to-end (logs + a small sample input → output artifact).

3. Teams Using LLMs

  • Containerized RAG pipeline with scripts for chunking, vectorization, and DB integration.
  • Documentation of the pipeline design and usage.
  • Logs showing pipeline runs with sample data.
  • Evidence of one end-to-end retrieval example.

4. Teams Using Vision or Other Modalities

  • Containerized pipeline for data ingestion and preprocessing.
  • Baseline model training/adaptation scripts with documentation of datasets, key hyperparameters, and model versions.
  • Experiment logs showing early results and planned next iteration.

5. Application Mock-up

  • Application mock-up/wireframe plus a minimal working app skeleton showing basic UI and back-end interaction.

Submission Instructions

All deliverables must be submitted via GitHub (milestone2 branch); submit the full commit hash on Canvas by 9:00 PM ET, October 20th.

TF Presentation & Evaluation

This milestone is evaluated through a 15-minute team presentation (with slides) to your assigned project TF, held during your regular weekly TF meeting.

  • Format: 15 minutes of team-led presentation, followed by Q&A from the TF.
  • Q&A scope: The TF will ask questions of the team and of individual members, covering both the slides and the details of your GitHub repository (code, configs, infrastructure, tests, design choices).
  • Shared accountability: Every team member must be prepared to answer questions on all parts of the project. “I didn’t do that part” is not an acceptable answer. Divide the labor however you choose — but the understanding must be shared across the entire team.
  • Slides: Bring slides that walk through what you built, why, and what’s next. Be ready to navigate to specific files in your repository when asked.