Vahdettin Karataş
Freelance data / ML delivery — batch workflows & internal APIs
  • Location:
    Prague, Czech Republic
Technical focus
  • Operational data to model-ready tables
  • Reusable transform definitions
  • Validation & structured errors
  • Internal APIs & automation
  • Pilot-sized engagements
Pilot capability · repeatable feature prep

Feature Store Mini

Turn an agreed raw export into a model-ready feature table with named transforms and checks—useful for pilots where spreadsheets and ad-hoc scripts drift apart.

Scoped pilot pattern—not SaaS or a full data platform. Outcome is a repeatable table with clear technical boundaries.

Agreed raw contract
Validated feature table
Batch + API parity
Pilot-ready handover

Limitations & scope

  • Not enterprise feature-store software: no Feast/Tecton-scale materialization, no real-time serving layer, and not a hosted multi-tenant product.
  • A single-repo implementation pattern for offline, batch features—not model training, scoring productization, or org-wide data governance.
  • Illustrative flows use Telco-style tabular data; your pilot pins the same idea to your own columns and approvals.

What this solves

Operational exports often arrive as messy spreadsheets: inconsistent columns, silent gaps, and one-off rules in different files. That makes batch scoring, internal analytics, and model refresh brittle—because nobody agrees what “clean input” means. This pattern locks a single definition of derived fields, runs one build, and surfaces validation in plain language before bad data reaches a dashboard or a training job.

What a pilot produces

From an agreed raw extract (CSV today), you get a repeatable feature table: stable column names, documented transforms, and a short validation summary (duplicates, schema checks, coarse sanity). Outputs are suitable for downstream modelling, internal reporting, or scheduled batch runs—without maintaining parallel “secret” notebook logic.

How the system works

Agree the raw contract
Named transforms in one build
Validate before hand-off
Same rules over HTTP when needed

Under the hood, definitions and checks live in code so batch jobs and a small HTTP API stay aligned. That reduces “it worked in my script” surprises when someone else runs the same data.

Scope and limits

This is intentionally not a consumer SaaS product, not a full enterprise Feast/Tecton-style platform, and not a substitute for company-wide data governance. It is a pilot-sized, implementation-level pattern: one pipeline, honest boundaries, and clear documentation. The live service is there so your technical stakeholders can verify behaviour—not the main business outcome of a successful pilot.

When this is useful

  • You need model-ready or decision-ready tables from recurring operational extracts without re-writing transforms each quarter.
  • You want schema and quality issues caught early, with summaries you can act on before reports or models consume the file.
  • You run batch scoring or internal analytics and need the same feature definitions every run, not ad-hoc spreadsheet logic.
  • You prefer a scoped engagement with a clear handover pattern rather than buying a broad “data platform.”

Where to try it

This overview lives on the VahdetLabs project page. The referenced API host runs the same pipeline for technical review. It is an engineering reference, not a consumer landing experience.

Feature Store Mini

Named transforms · validation · pilot scope — not a SaaS data platform.

© Vahdettin Karataş. All rights reserved.