LectureHSE · May 23, 2026
DataOps / MLOps for AI Business
How managers govern data, models, release flow and economics for production AI
/ DataOps / MLOps · HSE 2026
Slide contents
1. DataOps / MLOps for AI Business
How managers govern data, models, release flow and economics for production AI
2. Alexander Polomodov
Technical Director & Fellow, T-Technologies
Architecture and engineering R&D.
AI adoption in SDLC.
Focus: governed production AI.
3. Here DataOps/MLOps is a management loop
Releasing models safely on top of a data platform
We skip the basics — No data-platform recap.
We look as managers — Decisions, owners, loop control.
We count economics — Data, features, inference, review.
4. 01. Management frame
Governed path from data to decision
5. Manager Decisions
Experiment → pilot → production boundary.
Standard release path and checks.
Platform vs domain ownership.
Outcome governance metrics.
6. Operating-model options
Responsibility matters more than tool choice
Managed / self-service — Pipelines, registry, serving, observability; platform standards.
Domain ML teams — Quality, hypothesis, error cost and retrain stay with teams.
Hybrid — Shared standard path; dedicated runtime for critical scenarios.
7. Evaluate the loop, not the stack
What speeds up results
Hypothesis → live validation.
Repeatable release path.
Self-service without losing control.
What contains risk
Reproducible data, features, evals.
Stop rules, rollback, blast radius.
Unit cost and path reversibility.
8. 02. Data and feature contracts
Which data and features are safe for models
9. Data platform is not MLOps
Data platform
Data delivery, catalogs, lineage, access.
Reliable path to data products.
DataOps/MLOps
Training/inference/release data fitness.
Features, evals, release, runtime, feedback.
10. The model lifecycle is governed
Each handoff should leave an artifact: data, features, evaluation, release decision and rollback plan
11. Centralize feature contracts, not domain logic
A feature becomes a product contract: schema, owner, freshness, logic version and production readiness
12. Data antipatterns for ML
Unreproducible training slice.
Features without owner/version/SLA.
Offline metrics hide production skew.
Retrain hides data/policy debt.
13. 03. Safe model release
Model release must be governed
14. Release gates: replay, shadow, canary, A/B, rollback
Different release modes answer different questions: what changes, where risk sits and how fast rollback can happen
15. Who owns a model release
Decision participants
Product: error cost, outcome.
ML: quality, calibration, risk.
Platform: release path, rollback, SLO.
What breaks without a matrix
Quality separate from risk.
Release without consequence owner.
Rollback after degradation.
16. Model Quality Is Not the Decision
AUC misses cost, latency, review.
Threshold/policy/fallback change outcome.
Release card records what changed.
Decision combines quality, risk, cost.
17. 04. Serving and runtime economics
A production model costs money every minute: in features, inference, review and operational support
18. Online, batch and streaming inference
Choose the inference mode by latency, error cost, capacity cost and freshness requirements
19. Managed or owned runtime
When managed is enough
Fast launch, standard load profile.
Baseline SLO and monitoring are enough.
Error cost < platform complexity.
When to own
Custom routing, fallback, strict SLA.
Predictable volume; unit cost matters.
Special access/audit/environment policies.
20. FinOps for models: how to calculate cost
Features: backfill, materialization, freshness, storage.
Training/evals: experiments, retrain, baseline, replay.
Serving: CPU/GPU, peak, batch/online, cache.
Review: labeling, queues, escalation, corrections.
21. Cost must map to an owner
How to allocate cost
Scenario: antifraud, recommendations, support.
Model/endpoint: expensive routes.
Domain/product: decision owner.
How to use it in governance
Showback reveals behavior first.
Chargeback after metrics and rules.
Optimize architecture, not quarter budget.
22. 05. Observability, feedback and maturity
After release you need a system that sees degradation, collects feedback and scales without heroics
23. Human review and quality governance
Human-in-the-loop should be a controlled path for learning, review and escalation, not a permanent workaround
24. Maturity matrix: pilot -> growth -> regulation -> scale
DataOps/MLOps maturity shows up in repeatable release, risk ownership, observability and unit economics
25. Manager's summary matrix
Managed service: common path, fast launch.
Platform: release, observability, SLO, FinOps.
Domains: hypothesis, error cost, feedback.
Own runtime: control, scale, economics.
A production model is a business loop.
26. Links and materials
Google Cloud — MLOps pipelines.
Sculley et al. — Hidden Technical Debt.
Feast Feature Store documentation: docs.feast.dev
NIST AI Risk Management Framework 1.0: nist.gov/itl/ai-risk-management-framework