Capstone Project

For the shared project rules, see Final Project (Zoomcamp Logistics). For this course, follow the requirements here. The canonical version is in the projects folder of the repository.

Goal

Build a production multimodal agent and share it publicly. The capstone starts in Module 1 and has its first submission deadline on January 15, 2027.

Requirements

  • A multimodal agent that uses at least 2 GPU media models (any combination of image, video, and audio), with an agent that has 2 or more tools.
  • An async workload: the agent orchestrates long-running GPU jobs with job IDs, state persistence, and resume, not blocking calls.
  • Assets stored in object storage.
  • A batch run: 100 records for image and audio assets, 10 records for video assets.
  • Evaluation: a golden set plus a regression gate, at the system, agent, and media layers.
  • Observability: a trace per request that answers why this cost and why this long, plus metrics, logs, and a dashboard.
  • An SLO and a cost report.
  • Proof that a killed run resumes without redoing finished work.
  • Runs on Nebius Serverless AI (Jobs, Endpoints) as the reference execution layer.
  • The code is open source (MIT), OpenAI- and S3-compatible.

Bring your own records: a CSV or JSONL file with an id column. Records go in, generated media comes out.

Ideas: a marketing studio, podcast producer, video localization, storybook generator, product ad agent, or paper-to-explainer video.

Timeline

  • Module 1 (November 16 to 29): pick your records, get the first GPU generation working, and fill in the capstone proposal template.
  • Modules 2 to 4 (November 30 to December 20): teaching, with the capstone brief in Module 4.
  • Attempt 1: submit by January 15, 2027, and complete peer reviews by January 22, 2027.
  • Attempt 2: submit by January 29, 2027, and complete the final peer reviews by February 5, 2027.
  • Submission and peer-review deadlines are 23:59:59 UTC on these dates.

Peer review

Capstones are peer-reviewed against an 8-area rubric. Peers score each area 0, 1, or 2:

  • System: a working end-to-end system on Nebius Serverless AI with the required batch.
  • Agent: at least two tools, orchestrating long-running GPU jobs with job IDs and persisted state.
  • Media: at least two GPU media models, with assets stored in object storage.
  • Cost: a cost report that explains the cost per asset or per request.
  • Reliability: a killed run resumes without redoing finished work.
  • Observability: a trace per request, plus metrics, logs, and a dashboard.
  • Reproducibility: the run is reproducible from public MIT code, with pinned versions and inputs.
  • Presentation: the work is shared publicly with a clear write-up of what was built, how to run it, and the results.

To pass, submit your project, review peers, and share your work publicly. See the repository for the full score descriptions.

Certificate

A certificate is not confirmed yet. We will update this page when that is decided.