Skip to content
⁨NDDev AI⁩, home page
Community
Language: English
Display
Theme
Motion

MLOps: deploying and maintaining ML models

NDDev puts ML models into production and maintains them for data teams and products with AI features: we package models as services, version them in MLflow, check quality before every release, watch data drift, latency and cost, retrain on schedule and roll back on failure. My Attention AI, which we build as a technology partner, serves users today.

Who it is for

Data teams whose models are ready while release, monitoring and retraining are done by hand, and companies whose AI features already serve users and need looking after.

  • Model registry in MLflow
  • Quality checks before release
  • Drift and cost monitoring
  • Rollback to the previous version

What we set up around the model

We package the model as a service on FastAPI or vLLM and run it in a container with a pinned image version. Every model version sits in the MLflow registry with links to its training data, code and metrics. Releases go through CI: tests, a run on the control set and a comparison with the version in production.

  • packaging the model as a service and container
  • a registry of versions and experiments in MLflow
  • CI for models: tests and a control set
  • monitoring of drift, quality, latency and cost
  • tracing of LLM requests in Langfuse
  • a retraining pipeline on Airflow or Prefect
  • a release and rollback procedure
  • rules for storing and deleting data

After release we monitor input data, answer quality, latency and cost, and trace language model requests in Langfuse and OpenTelemetry. When the data drifts noticeably from the training data or a metric drops, an alert fires: the model retrains through a pipeline on Airflow or Prefect or rolls back to the previous version.

We agree separately on data: who stores prompts and answers, for how long and how they are deleted when the work ends. Service metrics are kept apart from user content, which is also how the telemetry of My Attention AI is built.

Taking a model into production

First we look at how the model lives today and fix the riskiest points.

  1. Audit

    We look at where and how the model is trained and run, who releases it and what happens when it fails.

  2. Service and registry

    We package the model and set up a version registry and reproducible training.

  3. Release pipeline

    We set up CI with tests and a check on the control set.

  4. Monitoring

    We connect metrics, alerts and dashboards and agree the thresholds.

  5. Retraining and support

    We launch retraining on a schedule or on a drift signal and take the system into support.

Work in this area

  • Platform

    My Attention AI: a platform for attention and organising your activity

    My Attention AI is a platform for working with attention and organising activity, with four areas: Screen, Together, Train and Care. The product is open to users and works in three languages, including English and Kazakh. It belongs to a separate company, My Attention AI, Inc., where NDDev's CEO Danil Silantyev is Chief Technology Officer.

Questions data teams ask about MLOps

We already have a model. Where do we start?

With an audit: we review the code, data, serving setup and metrics. You receive a list of risks and a plan for what to set up first.

Will you take on a model trained by another team?

Yes. We need access to the code, the data or a description of it, and the environment where the model runs. If training cannot be reproduced, restoring it comes first.

What is data drift?

Data changes over time: new products, different customers, seasonal swings. A model trained on old data starts to make more mistakes. Drift monitoring spots this before users see the errors.

Where will the model run?

On your servers, in an agreed cloud or on GPUs you rent. We set up the service, monitoring and retraining in the same place and hand documentation to your team.

How much does MLOps cost?

The price block gives the starting figure. The total is shaped by the number of models, how they run today and your monitoring requirements. Servers and computing are paid for separately. After launch the system can stay on support, usually for 10% of the project price per month.

from ₽280,000

Free warranty
40 days

Prices are starting prices for information only and are not a public offer. The final price is fixed in the specification.

Documents, video and forecasting

Tell us which models you already run

Describe your models, your data and how releases happen today. We will suggest a plan and give you a price.

We reply within 24 hours

Assistant

Loading the assistant…