Insights
Technical writing on computer vision, language AI, data science, MLOps and applied research. Grounded in practice, documented from real development work.
Most organisations have video infrastructure producing terabytes of footage that is reviewed only after incidents. We explore how computer-vision pipelines convert passive surveillance into structured operational data — from detection and tracking through zone analytics to real-time alerting.
Read articleA model that performs well in a Jupyter notebook often collapses under production conditions. We examine the six most common failure modes: data drift, missing preprocessing parity, latency assumptions, integration gaps, monitoring absence and ownership ambiguity.
Read articleThe deployment target changes everything — model format, latency budget, security model and cost structure. This article provides a decision framework for selecting between edge, on-premise and cloud inference.
Read articleMean average precision is a useful starting point, but operational deployment demands per-class analysis, confusion matrices, latency profiling, failure-case auditing and site-specific validation.
Read articleRAG systems fail silently. Retrieval quality degrades, hallucinations slip through, and users lose trust. We document our approach to chunking, embedding selection, retrieval evaluation and hallucination measurement.
Read articleDataset quality is treated as a labelling task. It should be treated as a software engineering discipline — with version control, automated validation, inter-annotator agreement and continuous integration for data.
Read article