[{"data":1,"prerenderedAt":158},["ShallowReactive",2],{"case-study-de-product-classification":3,"case-study-related-de-product-classification":135},{"id":4,"title":5,"card":6,"diagrams":12,"extension":24,"eyebrow":25,"facts":26,"intro":36,"layout":37,"locale":53,"meta":54,"order":52,"results":55,"slug":64,"stem":65,"storySections":66,"techGroups":73,"__hash__":134},"caseStudies\u002Fcase-studies\u002Fde\u002Fproduct-classification.json","MLOps auf AWS: ein Modell zur Produktklassifizierung trainieren und ausliefern",{"title":7,"badges":8,"thumb":11},"Produktklassifizierung mit MLOps",[9,10],"Amazon Bedrock","PyTorch","\u002Fcase-studies\u002Fproduct-classification\u002Ftraining-pipeline.webp",[13,16,20],{"src":11,"alt":14,"caption":15},"Trainingspipeline: von Produktdaten zu einem versionierten Modell","Die Trainingspipeline.",{"src":17,"alt":18,"caption":19},"\u002Fcase-studies\u002Fproduct-classification\u002Finference-pipeline.webp","Inferenzpipeline: von rohen Produktdaten zu strukturierten Ergebnissen","Die Inferenzpipeline.",{"src":21,"alt":22,"caption":23},"\u002Fcase-studies\u002Fproduct-classification\u002Fcore-stack.webp","Kern-Stack","","json","Case Study",[27,30,33],{"label":28,"value":29},"Branche","E-Commerce",{"label":31,"value":32},"Meine Rolle","Aufbau der Trainings- und Inferenzpipelines",{"label":34,"value":35},"Schwerpunkt","MLOps, Modelltraining","Ein großer Onlinehändler, ein Katalog mit Millionen Produkten und ein Modell, das rohe Produktdaten in einheitliche, strukturierte Informationen verwandelt.",[38,40,42,45,47,49,51],{"type":39},"header",{"type":41},"results",{"type":43,"index":44},"section",0,{"type":46,"index":44},"diagram",{"type":46,"index":48},1,{"type":50},"technology",{"type":46,"index":52},2,"de",{},[56,60],{"tag":57,"title":58,"body":59},"Traffic","Mehr Traffic durch bessere Auffindbarkeit","Einheitlich klassifizierte Produkte sind in Suche und Navigation leichter zu finden.",{"tag":61,"title":62,"body":63},"Umsatz","Mehr Umsatz durch bessere Produktdaten","Produkte, die häufiger gefunden werden, verkaufen sich häufiger.","product-classification","case-studies\u002Fde\u002Fproduct-classification",[67],{"heading":68,"paragraphs":69},"Die Geschichte",[70,71,72],"In einem Katalog mit Millionen Produkten kommen Produktdaten unvollständig und uneinheitlich an. Von Hand zu klassifizieren skaliert nicht, das Ziel war also ein trainiertes Modell, das es automatisch tut, auf Basis unterschiedlicher Arten von Produktdaten.","Ich habe die Trainingspipeline gebaut. Python-Jobs lesen Produktdaten aus DynamoDB und RDS, bereiten Features auf, erzeugen Text-Embeddings mit Amazon Bedrock und trainieren das Modell. Daten und versionierte Modellartefakte liegen in S3.","Ich habe auch die Inferenzpipeline gebaut, einen Python-Service auf ECS, der neue und geänderte Produkte klassifiziert und die Ergebnisse zurück in den Katalog schreibt. Überwacht wird er in CloudWatch, ausgeliefert über GitHub CI\u002FCD.",[74,104,123],{"title":75,"items":76},"AWS",[77,80,84,88,92,96,100],{"name":9,"role":78,"icon":79},"Text-Embeddings für Produktdaten","\u002Ficons\u002Faws\u002FAmazon-Bedrock.svg",{"name":81,"role":82,"icon":83},"Amazon ECS","Feature-Jobs, Inferenz-Service","\u002Ficons\u002Faws\u002FAmazon-Elastic-Container-Service.svg",{"name":85,"role":86,"icon":87},"Amazon EC2","Modelltraining","\u002Ficons\u002Faws\u002FAmazon-EC2.svg",{"name":89,"role":90,"icon":91},"Amazon S3","Trainingsdaten, Modellartefakte","\u002Ficons\u002Faws\u002FAmazon-S3.svg",{"name":93,"role":94,"icon":95},"Amazon DynamoDB","Produktdaten","\u002Ficons\u002Faws\u002FAmazon-DynamoDB.svg",{"name":97,"role":98,"icon":99},"Amazon RDS","relationale Produktdaten","\u002Ficons\u002Faws\u002FAmazon-RDS.svg",{"name":101,"role":102,"icon":103},"Amazon CloudWatch","Metriken, Logs, Alarme","\u002Ficons\u002Faws\u002FAmazon-CloudWatch.svg",{"title":105,"items":106},"Machine Learning",[107,111,113,117],{"name":108,"role":109,"icon":110},"Python","Pipelines und Service","\u002Ficons\u002Ftech\u002Fpython.svg",{"name":10,"role":86,"icon":112},"\u002Ficons\u002Ftech\u002Fpytorch.svg",{"name":114,"role":115,"icon":116},"scikit-learn","Baselines und Evaluation","\u002Ficons\u002Ftech\u002Fscikitlearn.svg",{"name":118,"role":119,"icon":120,"icons":121},"pandas, NumPy","Feature-Aufbereitung","\u002Ficons\u002Ftech\u002Fpandas.svg",[120,122],"\u002Ficons\u002Ftech\u002Fnumpy.svg",{"title":124,"items":125},"Delivery",[126,130],{"name":127,"role":128,"icon":129},"GitHub Actions","CI\u002FCD für Pipelines und Service","\u002Ficons\u002Ftech\u002Fgithubactions.svg",{"name":131,"role":132,"icon":133},"Docker","Container","\u002Ficons\u002Ftech\u002Fdocker.svg","v5wQQPeHvvg_vGZNENLBWyRYj5oeDHoH1FlV85108Cw",[136,143,150],{"slug":137,"card":138},"property-management-platform",{"title":139,"badges":140,"thumb":142},"Mandantenfähige Hausverwaltungs-Plattform",[141,9],"Mandantenfähigkeit","\u002Fcase-studies\u002Fproperty-management-platform\u002Fsystem-architecture.webp",{"slug":144,"card":145},"ale-business-bookkeeping",{"title":146,"badges":147,"thumb":149},"Buchhaltung per Chat",[148,9],"LLM-Harness","\u002Fcase-studies\u002Fale-business-bookkeeping\u002Fsystem-architecture.webp",{"slug":151,"card":152},"kv-cache-storage-tiering",{"title":153,"badges":154,"thumb":157},"KV-Cache-Storage-Tiering",[155,156],"vLLM","LMCache","\u002Fcase-studies\u002Fkv-cache-storage-tiering\u002Ftier-hierarchy.webp",1790607906452]