Before you tune the AI model, inspect the data
Model changes are easy to see. In practice, inconsistent labels, weak provenance, and missing domain definitions are often the real ceiling.
Read insightWORK & INSIGHTS
The work, the reasoning, and the evidence.
Client projects, independent builds, and research notes—each with the problem, the approach, what was verified, and what was not.
Extracting age from an Instagram bio is a context problem
A real client project started with a deceptively simple request: recognize “25 yo” or a birth year without treating “married in 2018” as someone's age.
Read insight
What medical AI work taught me about domain expertise
A research collaboration with physicians explored deep learning in breast cancer histopathology. The lasting lesson was how much the model depends on clinical definitions, data provenance, and honest evaluation.
Read insight
What my UAV master’s thesis taught me about data quality
My 2021 thesis combined 2D and 3D segmentation, scene comparison, geocoding and point clouds. Its most durable lesson is that model quality depends on the assumptions hidden in the data.
Read insight
Automating a spreadsheet that came back every month
A client had an Excel file that was updated every month and had to match a fixed target format. The useful part was turning that visual target into repeatable rules.
Read insightAnomaly Detection Learning Platform: start with the use case
A real client asked for a clear Google Colab demo of several anomaly detection algorithms. The hard part was not calling fit(); it was making the data and the result understandable.
Read insight
Large Dataset Filtering with Python
A public project built around a practical question: how do you filter a CSV that is too large for a normal spreadsheet without hiding the rules from the person using it?
Read insightEarlier technical writing is also available on Mehmed's Medium profile.