The interesting failures in data products often happen before the model. These notes cover ingestion contracts, evolving schemas, PostgreSQL, retries, queues, multilingual source data, data quality, and the decisions required to keep pipelines observable and recoverable.
How I structure workspace isolation, permissions, background jobs, billing and audit state when turning a FastAPI application into a multi-tenant SaaS.
The data-model and reliability problems that shaped TenderWise's TED ingestion pipeline, from notice versions and lots to multilingual fields, retries and semi-structured qualification criteria.
Why late-payment prediction needs more than a chronological train/test split, and how I reconstruct what was actually knowable when each Payrithm prediction would have been made.
What running PostgreSQL, Redis, Celery workers and multiple SaaS applications on an 8 GB VPS taught me about resource budgets, concurrency, queues and operational simplicity.
How I designed TenderWise as a multilingual procurement SaaS that ingests TED notices, applies deterministic profile matching, separates relevance from qualification, and turns tender analysis into collaborative bid preparation.
The often-overlooked craft of transforming raw data into informative features — covering encoding, scaling, missing values, interactions, and selection.
How to build a robust anomaly detection pipeline for multivariate sensor data using LSTM autoencoders in PyTorch, with practical tips on thresholding and deployment.
I use Google Analytics to understand how visitors use this portfolio and improve it. Analytics cookies and page-view tracking remain disabled unless you accept. You can change your choice later from the footer.