Build, scale and monitor agents that run GPU image, video and audio models on Nebius Serverless AI. Free, open-source (MIT), project-based.
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: a Santa gift-card run and a non-seasonal product catalog to ad creative run.
- 4 modules over 5 weeks (Module 1 is two weeks) · ~8-10 h/week · free · open-source · project-based
- Week 1: Make it work
- Week 2: Make it intelligent
- Week 3: Make it scale
- Week 4: Make it production-ready
- 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
Build consistent media generation on serverless GPUs.
- Landscape: modalities, model families, size/VRAM/latency/quality, 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, 7 metrics)
Orchestrate multi-agent media generation workflows.
- 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)
Scale batch media generation.
- 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 retry / timeout / fallback / checkpoint / resume / idempotency / dead-letter; kill → resume proof
- 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)
Evaluate, trace and set SLOs for media generation agents.
- Evaluating generative systems: system / agent / media layers; judges, deterministic checks, human eval; golden set + regression gate
- Observability: one trace, "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)
A production multimodal agent that uses at least 2 GPU media models, peer-reviewed. It starts in Module 1 and is due 15 Jan 2027. Share it publicly and run it on Nebius Serverless AI.
- 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 per-call media APIs who want to know when self-hosting wins
Medium-advanced level. Application only, limited seats - tell us what you are building.
Python, Docker basics, HTTP APIs, command line. No diffusion theory required. Prior LLM or MLOps course experience helps, not required.
MIT license. Code is OpenAI- and S3-compatible. Nebius Serverless AI (Jobs, Endpoints) and Nebius Token Factory are the reference execution layer.
Application only, limited seats. Tell us what you are building. We review every application, usually within 2 business days. Approved applicants get the full path (homework, capstone, peer review) and Nebius Builders Program credits. Everyone can follow the public repo.
- Welcome session: Tue 10 Nov 2026
- Course starts: Mon 17 Nov 2026 · apply by Wed 26 Nov 2026
- Module 1 (two weeks): 17-30 Nov 2026 (through Sun 30 Nov)
- Module 2: 1-7 Dec 2026
- Module 3: 8-14 Dec 2026
- Module 4: 15-21 Dec 2026
- Course ends: Sun 21 Dec 2026
- Capstone due: Fri 15 Jan 2027
- Nebius DevRel team (course content and delivery)
- Alexey Grigorev (DataTalksClub - course logistics and community)
DataTalks.Club is a community of data enthusiasts learning and growing together. We're all about sharing knowledge, helping each other out, and making data science more accessible.
Join us: • Website • Slack Community • Newsletter • Events • Calendar • YouTube • GitHub • LinkedIn • Twitter