Rabia Yoruk

AutoML platform for a telecom company

An AutoML platform for one of Türkiye's largest telecommunications companies. Its data flow module works like KNIME: each data step is a separate node on a canvas. I led the build as tech lead.

Role
Tech Lead
Where
OptiWisdom
When
2023 to 2025
Stack
  • Next.js
  • Python services
  • pandas
  • Node graph runner
  • AWS
  • GCP

20+

node types in the data flow editor

5

client projects running in parallel, on average

20+

client projects delivered at OptiWisdom

Data flow module, simplified. Each step is its own node. Client details are not public. Flow: Uploaded datasets and Database tables, then Pre-operations, then Post-operations, then Flow runner, then Generated dataset and Export to DB, then AutoML experiment. Uploaded datasets files Database tables connections Pre-operations merge, pivot Post-operations filter, fill NA Flow runner node by node Generated dataset 100-row preview Export to DB table, schema AutoML experiment train models Data flow module, simplified. Each step is its own node. Client details are not public. Flow: Uploaded datasets and Database tables, then Pre-operations, then Post-operations, then Flow runner, then Generated dataset and Export to DB, then AutoML experiment. Uploaded datasets files Database tables connections Pre-operations merge, pivot Post-operations filter, fill NA Flow runner node by node Generated dataset 100-row preview Export to DB table, schema AutoML experiment train models
Fig. 1 Data flow module, simplified. Each step is its own node. Client details are not public.

Problem

Before a model can be trained, the data has to be joined, reshaped and cleaned. The platform lets teams do these steps visually, one node at a time, and then train models in the same place, without writing code for each step.

How the data flow works

  • Users drag datasets and operation nodes onto a canvas, like in KNIME.
  • Each node is a separate step with its own parameters. Pre-operations reshape tables: merge, concat, aggregate, pivot, melt, split, transpose, or a Python script. Post-operations clean columns: filter, sort, rename, fill NA, drop duplicates, type cast, resample, clip.
  • The canvas is saved as a graph of nodes and edges. On run, the backend executes it node by node. A node turns yellow while it runs, green when it is done and red when it fails, and the log shows the error.
  • The result is a new dataset with a 100-row preview. It goes into an AutoML experiment, or it is exported to a database table.

What I did

  • Led the build of the platform: Next.js front end and Python services.
  • Owned architecture and stack decisions for AI, data and software products at OptiWisdom.
  • Set up and ran client projects on AWS and GCP, and built event driven services with Kafka and Node.js.
  • Ran hiring, agile delivery and performance tracking with the founder.