Delivering AI and analytics products end to end for 5 enterprise clients in a 5-person squad (data scientists, engineers, product): problem definition, scoping, delivery and measurement. Ran product discovery for a national retail chain, defined 7 AI use-cases — all entered development — and put a working demo in the client's hands within 2 weeks. For a national logistics operator: demo in 3 days, then an image-analysis product covering 500 in-vehicle cameras continuously, replacing an entirely manual process. Led the model selection decision on a TÜBİTAK-supported domain-specific marketing LLM: evaluated 4 Qwen 3.5 variants on a purpose-built 500-question benchmark scored with an LLM-as-judge method; the 27B model did not win. Identified 12 metric inconsistencies in the evaluation pipeline and rebuilt it. Designed a 200,000+ record synthetic data pipeline; training on H200 GPUs.