Production AI Agents for Generative Media

Build, scale and monitor agents that run image, video and audio models on serverless GPUs.
Free · open-source (MIT) · project-based · by Nebius × DataTalks.Club

Starts Mon 16 Nov 2026 · Apply by Thu 12 Nov 2026 · Apply here

How do you build a production AI system when an agent can launch expensive, slow, failure-prone GPU workloads that produce large media artifacts?

In this course, you build one production system that handles it: a Personalized Media Factory. A record (customer, product, user) goes in. An agent writes the copy, generates an image plus a voice or video asset, evaluates the result, regenerates on failure, and scales from 1 to 1,000+ records with a cost line per asset.

Two datasets ship with the course:

• Santa gift cards: personalized gift cards generated for a list of children
• Product catalog: ad creative generated for a non-seasonal product catalog

Format: 4 modules over 5 weeks (Module 1 runs for two weeks), ~8–10 h/week, plus a capstone project.

1. Make it work
2. Make it intelligent
3. Make it scale
4. Make it production-ready

What You'll Learn

• How to run image, video and audio models as Jobs and Endpoints on serverless GPUs, without managing a cluster
• How to give agents GPU models as tools (MCP) and build generate → evaluate → regenerate loops
• How to fan out batch workloads on spot instances, handle failures and report cost per asset
• How to add evals, tracing and SLOs so you can say what a media agent costs and when it is good enough to ship

Modules

Module 1: Generative Media Workloads

Get media generation running reliably on serverless GPUs, from a local script to a containerized GPU worker.

• Landscape: modalities, model families, size/VRAM/latency/quality trade-offs, why video is expensive, licences
• First GPU generation: local script → container → GPU Job → Object Storage
• Containers for generative AI: CUDA/PyTorch images, weights vs image, HF cache, pinned revisions, secrets, I/O
• GPU engineering lab: precision, quantization, attention backends, offloading, VAE tiling, batching, torch.compile, caching, GPU selection, OOM debugging
• Jobs vs Endpoints
• Homework 1: GPU Media Worker (image + audio or video, measured on 7 metrics)

Module 2: Agents That Use GPUs

Build an agent that plans the work, calls GPU models as tools and handles long-running jobs.

• The agent loop, no framework
• Tools and structured outputs: schemas, validation, retries, idempotency, timeouts, errors, state; tools as an MCP server
• CPU agents + GPU tools: control plane vs compute/data plane
• Long-running tools: async, polling vs events, callbacks, job IDs, state persistence, resume, duplicate execution, job lifecycle
• Agent vs workflow; hybrid architectures
• Homework 2: Media Agent (coffee ad: plan → generate → inspect → regenerate → video → narration → voice → assemble → trace)

Module 3: Scaling, Reliability and Cost

Go from one request to 1,000, survive failures and know the cost of every asset.

• One request to 1,000: concurrency, queues, backpressure, rate limiting, autoscaling, scale-to-zero, batch vs real-time, jobs vs endpoints, GPU capacity, heterogeneous pools; manifest → Jobs fan-out → fan-in
• Failure engineering: break 11 things, then add retry / timeout / fallback / checkpoint / resume / idempotency / dead-letter; prove kill → resume works
• Cost engineering: cost-per-request model, cost-aware agent (cheap draft → evaluate → expensive)
• Performance optimization: quality/latency/cost frontier, benchmark sweeps
• Homework 3: Santa Scale Challenge – Media Edition (100 → 1,000 children, leaderboard)

Module 4: Evaluation, Observability and Production

Measure quality, trace every run and decide when the system is ready to ship.

• Evaluating generative systems: system / agent / media layers; judges, deterministic checks, human eval; golden set + regression gate
• Observability: one trace that answers "why $1.83 / why 94 s"; traces + metrics + logs, dashboards
• Production architecture: auth, secrets, quotas, abuse, prompt injection, content safety, isolation, retention, licensing, provenance, reproducibility, release (versioning, canary, rollback)
• Homework 4: SLO on the Homework 3 run (evals, trace, dashboard, alert, go/no-go)

Capstone Project

A production multimodal agent that uses at least 2 GPU media models, peer-reviewed. You start it in Module 1 and submit it by Fri 15 Jan 2027. Share it publicly and run it on Nebius Serverless AI.

Who Should Join

• AI/ML engineers adding image, video or audio generation to real systems
• MLOps and platform engineers who own the batch pipeline, scheduling and cost
• Founders and small teams paying for per-call media APIs who want to know when self-hosting wins

Level: intermediate to advanced.

Prerequisites

• Python
• Docker basics
• HTTP APIs
• Command line

No diffusion theory required. Prior LLM or MLOps course experience helps but isn't required.

Dates

How to Join

The course is free. Seats on the full track are limited, so it is application-only.

1. Apply here and tell us what you are building.
2. We review every application, usually within 2 business days.
3. Approved applicants get the full track: homework, capstone, peer review and Nebius Builders Program credits.

Not applying? All materials are in this repo and anyone can follow along. You'll need your own GPU access to run the code.

Cost

The course itself is free. The workloads run on GPUs, which cost money:

• Approved applicants: Nebius Builders Program credits cover the course workloads.
• Following on your own: you pay for your own GPU usage. Each module reports cost per asset, so you'll know what a run costs before you scale it.

Open and Portable

• License: MIT
• Portable code: works with OpenAI-compatible APIs and any S3-compatible object storage
• Reference execution layer: Nebius Serverless AI (Jobs, Endpoints) and Nebius Token Factory

Your Instructors

• Nebius DevRel team (course content and delivery)
• Alexey Grigorev (DataTalksClub - course logistics and community)

About DataTalks.Club

DataTalks.Club is a community of data enthusiasts who learn and grow together. We share knowledge, help each other, and make data science more accessible.

Join us:
• Website
• Slack Community
• Newsletter
• Events
• Calendar
• YouTube
• GitHub
• LinkedIn
• Twitter