The notebook model has AUC 0.87. In production the credit application waits 14 days, because the “30-day transaction count” feature is a manual GUI export to CSV. Or inference p99 jumps above 500 ms, because someone pickled a laptop artefact and stuffed it into a container with no Feature Store.
This is not an algorithm problem. It is a pipeline problem. Enterprise predictive analytics dies on a hybrid: Jupyter and Dataiku on one side, Spark CLI, Airflow and kubectl on the other. The team clicks, pastes, exports. The model ages at the speed of the “to deploy” backlog.
What most “we need a Data Scientist” briefs miss: when to rent one person (data science staff augmentation, body leasing) and when the pipeline needs a squad of DE + DS + MLOps. Hourly rates are in a separate piece: IT body leasing rates in Poland 2026. Here we count why AUC alone never reaches production.
One sentence that never makes the sales slide
A Data Scientist without a data pipeline is a researcher with an 8-hour Jira cap. A pipeline without a Data Scientist is ETL that predicts nothing. Staff-augmenting one role works when the other already sits in your team. Otherwise you bought a timesheet, not a prediction.
1. Anatomy: a notebook is not a product
A typical day on a predictive-analytics team in a bank, telco or retailer does not look like a scikit-learn tutorial. It looks like switching modalities:
- GUI: quality checks in a BI tool, a manual warehouse export, clicking retrain in Dataiku / SageMaker Studio, watching drift on a dashboard.
- CLI / API: Spark-submit, dbt run, airflow dags trigger, kubectl rollout, MLflow log, a BigQuery batch.
Every switch is where reproducibility dies. A feature computed in a notebook is not the feature computed in the service. That is training-serving skew: the model learns a definition production will not repeat. Sculley et al. named this in 2015 pipeline jungles and glue code — ML technical debt that grows faster than the model itself.
Three signs you have a jungle, not a platform:
- CSV as the interface. The source of truth is yesterday’s export. Nobody can say which table and which filter.
- The feature is computed twice. Pandas on a laptop, SQL in Airflow. A 0.3 pp fill-rate gap shows up a quarter later, when drift is already inside the credit decision.
- Deployment as a ticket to DevOps. The Data Scientist stops at “model.pkl”. Someone else packs the image, a third person sets the probe. Idea-to-inference: weeks, not days.
That is why a “hire a Data Scientist” brief with no pipeline context is the most expensive way to get another notebook. The skills that actually ship predictive analytics sit on three landings: Data Science and BI, Data Engineering, MLOps.
2. What the papers say, not the MLOps decks
You do not need another “MLOps maturity” definition. You need the boundaries at which a model never leaves the lab.
- ML debt lives in the glue, not the loss. Sculley et al., Hidden Technical Debt in Machine Learning Systems (NIPS 2015): in ML systems the learning code is usually a small fraction. The rest is configuration, collection, verification, serving, monitoring. “Change one feature” detonates a cascade, because the dependencies are hidden. A pipeline jungle grows when you add paths instead of deleting old ones.
- Computer work is hybrid. Shi, Wang, Fang, Liang et al., CUA-Universe (arXiv:2609.05374, September 2026): real work mixes visual-state inspection with precise CLI over shared application state. GUI-only agents produce inefficient trajectories; CLI-only agents go blind on layout. Training on hybrid GUI+CLI moved a 9B model: CUA-Verse +39.3 pts, −37% steps, −60% tokens; OSWorld SR +16.8 pts, −57% steps, −44% tokens. This is not a paper about credit scoring. It is a measurement of what window-and-terminal juggling costs. A Data Science team that clicks Studio and pastes commands into Confluence pays the same tax.
- Delivery is a team metric, not a notebook metric. DORA 2024: elite vs low is an order of magnitude in deploy frequency and change lead time. Those numbers describe teams with a shared Definition of Done. Five separate timesheets (DS from vendor A, DE from B, “someone for Kubernetes” from C) will not compose a lead time. They will compose three replacement SLAs and one “in progress” status. We unpacked that failure mode in team leasing vs staff augmentation 2026.
The architectural takeaway
Standardising the pipeline is not “buy SageMaker”. It is killing GUI↔CLI switching wherever an API and IaC can replace it. Feature Store, one DAG, one serving image, a schema test on the PR. The rest is vendor cosmetics. If your Data Scientist cannot reproduce a feature from a commit, you do not have predictive analytics. You have a demo.
3. Production case: scoring that lagged a sprint
From engineering practice: 3 weeks down to 3 days
A financial institution, a book of millions of clients. Application scoring and fraud detection. Sources: Oracle / PostgreSQL, a data lake (S3 + Hive), logs in Elasticsearch. The predictive-analytics team worked like this:
- extraction: hand-written SQL, CSV export,
- features: Pandas on a laptop, Excel joins “just in case”,
- training: Dataiku GUI, manual artefact export,
- release: a ticket to DevOps, Docker “as it comes”, API Gateway, Kubernetes with no schema test.
Problem: idea-to-inference 3–4 weeks. Models were stale on go-live day. p99 of critical predictions > 500 ms. Run cost grew with every new pickle.
Change: Spark pipelines orchestrated by Airflow into a Feature Store (Hopsworks). Every stage (data, train, validate, serve) in an image. GitOps onto the cluster. Feature-drift and data-quality monitoring (Prometheus / Grafana). DS stops exporting CSV. DE stops guessing which feature version to take into the batch.
Measured after: time-to-production 3 days instead of 3 weeks. Pipeline and deploy errors −85%. Inference p99 < 80 ms. MLOps infra cost −40% — not from a cheaper cloud, from no manual firefighting and abandoned retrains.
That result did not come from “a better XGBoost”. It came from leaving the GUI for inspection and the CLI/API for the path that has to repeat at 03:00. Same lesson CUA-Universe measures on agents: hybrid works when both modalities share state — not when a human is the bus between windows.
4. Decision table: which pipeline, which roster
| Approach | Complexity | Latency p95/p99 | Infra cost | Team overhead | When to use |
|---|---|---|---|---|---|
| Notebook + ad-hoc ETL | Low at the start, grows exponentially | Seconds–minutes, often manual | Low at the start | High (glue, debugging) | PoC, one analyst, no SLA |
| MLOps OSS (Airflow, MLflow, Kubeflow, Feast/Hopsworks) | High (platform) | ms–s, repeatable | Medium | Medium–high (DE + MLOps) | Enterprise that needs control and on-prem / hybrid |
| Managed service (SageMaker, Vertex, Azure ML) | Medium (cloud integration) | ms–s | Medium–high (service meters) | Lower operationally | When you already live in one cloud IAM and are not hauling a legacy GUI |
| Hybrid + T&M roster | Very high | Bound by the weakest connector | High | High until GUI and API are stitched | Bank, pharma, data that does not leave. Usually a squad, not one role. |
Python CRUD and Python for data pipelines are two jobs. Backend bands 145–200 / 200–280 PLN/h (contractor / client). Data pipelines sit in the upper part of that band. Production AI/LLM: 180–250 / 240–350 PLN/h. Margin 10–25%. A map, not a price list — detail in the 2026 rates article.
5. Anti-patterns you will not find in a Kubeflow tutorial
- A “Python Developer” brief for scoring. You will get Django. You will not get a Feature Store. The market split: classic backend has supply, pipeline and LLM engineers do not. A mislabelled ad fills with CVs in 48 h and zero drift competence.
- No data tests. CI lints the model, not the input schema. A silent null change in the source ships. The prediction is “green”. The business decision is not.
- A feature with no owner. DS computes it in a notebook, DE in dbt, BI in Power Query. Three churn definitions. One dashboard, three stand-up wars.
- A fake analytics team. DS from vendor A, DE from B, MLOps “when we have time” from internal DevOps. Three onboardings, no shared DAG. That is not staff augmentation. It is the integration tax described under IT team leasing.
6. Playbook: who to rent, and in what order
Do not start with the model. Start with which modality is blocking the SLA.
- One gap in an existing pipeline. You have Airflow, a Feature Store and someone who reviews PRs. You are missing a senior on the model or on Spark. That is classic data science staff augmentation or a Data Engineer: one person, your stand-up, your DoD. First profiles in days, not a quarter — we source on demand, we do not sell a named bench for tomorrow morning.
- The notebook is the product. No DAG, no schema test, no serving. One person will not stitch that. Roster 3–5: DE (sources, dbt/Spark), DS (model, validation), MLOps (image, drift monitoring). Closer to team leasing than “we’ll add one more analyst”.
- Data does not leave the VPC. Contractor in your IAM, your cloud, NDA and processing terms. Colab with a production dump is a leak, not a speed-up. Offline feature computation in batch, online serving from the same feature code.
- Ramp-up. One person into an existing team: first CVs in 24–48 h, start after your interviews and contract. A squad from zero: weeks, not a sprint, because you are stitching access, data and DoD. “Three senior Python from Monday” is a CV pack or a bench we do not have — and will not pretend to have.
Commoditech has done T&M and perm search from Warsaw since 2012. 80+ specialists in network, not an idle bench. A T&M or success-fee brief can be filed from an IDE via MCP for AI agents, not only a contact form. A human quotes a stack; ranges live in the rates article, not in the agent’s JSON.
FAQ
How is data science staff augmentation different from hiring a Data Scientist?
A perm hire is an FTE in your HR. Average IT-ad lifetime on JobHunt in September 2026 is 50 days — before you even interview. Data science staff augmentation is T&M: the person in your team, your Git, your SLA, with a role-replacement clause. You are not buying “a model on subscription”. You are buying an hour of skill. TCO vs payroll (recruiting, bench, severance, hardware) is in the 2026 rates.
When should we rent one person vs a Data Science squad?
One person when the pipeline already exists and you are missing a gap: senior DS, Spark, MLOps. A 3–5 person squad when the notebook is the only artefact. A Data Scientist alone will deliver AUC. They will not deliver p99, a schema test or a 03:00 retrain. If nobody on your side can review the contractor’s PR, do not buy staff augmentation. Buy a lead plus one role, or a squad.
How much does data science staff augmentation cost in Poland in 2026?
Python for data pipelines is not CRUD: in practice the client pays the top of the 200–280 PLN/h senior band. When the model ships with LLM / RAG, closer to 240–350 PLN/h. Vendor margin 10–25% — if someone promises 8% with 5-day replacement, it is sitting on another line. A market map, not a quote. A human prices the brief.
Can a Data Science contractor work on NDA and GDPR data?
Yes: your environment, your IAM, your cloud or on-prem, a contract with NDA, processing terms and IP on the client side. A notebook on a private laptop with a production dump is a leak. Offline features in batch, online serving from the same code. Access audit is yours. The contractor does not take the dataset “to compute faster”.
Sources
- Sculley, D. et al. (2015). Hidden Technical Debt in Machine Learning Systems. NIPS 2015. papers.nips.cc
- Shi, H., Wang, W., Fang, W., Liang, Y. et al. (2026). CUA-Universe: A Scalable and Dynamic Environment for Hybrid GUI+CLI Agents. arXiv:2609.05374. arxiv.org/abs/2609.05374
- DORA / Google Cloud (2024). Accelerate State of DevOps Report.
- JobHunt.pl (1 September 2026). 21,211 active IT roles, 50-day average ad lifetime — time-to-hire vs T&M context.