Francisco
Perez-Sorrosal

Dataset Quality Improvement

Entity Reconciliation Model Training: Fine-tuned transformer-based language models on heterogeneously labeled datasets, integrating editorially curated data with automatically labeled data through entity linking via common identifiers. This approach aimed to deduplicate and reconcile meta-entities related to People and Creative Works.

Entity Reconciliation Pipeline: Developed an automated pipeline for training entity reconciliation models for YK, from data curation to inference testing.

Entity Reconciliation Inference

Kamino Project - Mail Classification System: Developed a deep learning pipeline using Hugging Face to train/evaluate models based on the Teacher/Student knowledge distillation framework. Starting with limited human-labeled training data, a large and complex teacher model was created to achieve high accuracy. This model was then used to generate teacher-labeled data for training lightweight student models optimized for deployment. Two types of student models were produced: (1) online models, optimized for speed and suitable for real-time classification tasks, and (2) offline models, designed to include richer information for improved accuracy in non-real-time applications.

AI/ML Working Group - AI/ML Future Strategy Definition in the Cloud: As Yahoo transitioned its infrastructure from on-premises data centers to a fully cloud-based environment, scientists and research engineers required modern tools to streamline their workflows. These tools needed to facilitate dataset and model sharing, ensure reproducibility of experiments, enable rapid experimentation and iteration on new models, and support seamless publishing and deployment into pre-production/production environments.

Contextual Targeting Solution: In the Ads platform, as we were moving towards a cookie-less world, the ability to track users' online signals for behavioral targeting would be drastically reduced, making contextual targeting an appealing alternative for advertising platforms. Using our experience in hierarchical multilabel classification in other contexts, we helped the Ads team to use it in category-based contextual targeting at Yahoo. We proposed and implemented a multilingual model that can accurately classify web pages into a hierarchical taxonomy (specifically, the Yahoo Interest Categories taxonomy) without crawling their content.

Deep Learning-based Multi-label Classification

Scalable Few-shot Classification with Parallel Prefix Conditioning: Class representations are encoded in parallel but produce independent binary labels for each input through a shared classification output layer.

Clickbait Classifier based on Transformers

Hierarchical Transfer Learning for Multi-label Text Classification

DL-based Multi-class Classification for Content Ingestion: Experimented with LSTM/GRU-based models for multi-class classification using TF/Keras. Explored transfer learning techniques for multi-label text classification.

Machine Learning Content Classification Pipeline: Built a pipeline for easy data ingestion, model training and evaluation based on SVMs for the Sieve platform.

ML Pipeline Scalability (k8s-make): Built k8s-make, a tool and a workflow-based framework to harness the compute-power in Yahoo's on-prem clusters to deploy a tailor-made Kubernetes cluster to parallelize the training of our SVM-based pipeline in a simple way.

Twitter Firehose: At-scale Ingestion Streaming System for Tweets: Reimplemented the new Twitter API (v2.0) in the firehose. Re-architected previous solution to be more performant yet keeping backwards compatibility. Collaborated with the Sports team for the productization of the new firehose.

Scalable Content Ingestion Platform (Omid OSS): Transferred Omid as full open-source project into the Apache Software Foundation (ASF). Continued supporting the Omid transaction manager project in production.

Omid presented at Hadoop Summit, 2016, San Jose, CA (USA).. Paper: 'Omid, reloaded: scalable and Highly-Available transaction processing' (USENIX FAST 2017, 20 Citations)..

Scalable Content Ingestion Platform (Omid HA): Worked at the multi-tenant content ingestion platform at Search organization. Added High Availability to the Omid Transaction Manager for HBase. Scaled Omid in multi-core architectures. Supported the Omid transaction manager in the production infrastructure of the content ingestion platform. Started exploring the transfer of Omid as open-source project to the Apache Software Foundation.

Omid - Transaction Manager for Big Datastores: Re-architected the original codebase and implemented new features. Supported the Omid integration in Sieve project (Yahoo's large scale Internet content ingestion platform). Presented Omid and its poster at Yahoo's internal technical conference (Techpulse Conf. 2014).

Edentity/Pachiderm - Ubiquitous Content Access and Management: Ubiquitous content access and management of personal data (images, video) held on 3rd party services (Flickr, GDrive, Dropbox). Designed and implemented a scalable synchronization data module. Defined and implemented a RESTful API for each module defined (User Registration, Search, Content Fetching).

RiddlR - Percolator-like Incremental Processing System: Prototyped the framework and built a multi-stage example application on top of it. Presented RiddlR and its poster at Yahoo's Techpulse Conf. 2013.

CumuloNimbo - 7th European Framework Programme (FP7-257993): Designed and implemented a prototype of an incremental processing system for Big Datastores. Added durability guarantees to HBase through BookKeeper. Represented Yahoo in project meetings and evaluation sessions.

Giddra Project - SONY Socialife Application: High Scalable Big-Data Backend Platform for web and social content ingestion and aggregation. Contributed to the architectural definition of the platform. Did coordination over multiple teams. Defined the REST API for allowing clients to access/use the backend.

NEXOF-RA: NESSI Open Framework - Reference Architecture: 7th European Framework Programme (FP7-216446). Contributed to a reference architecture (RA) for a European service platform. Built the specification of a set of architectural patterns for non-functional attributes and analysis of how to integrate cloud platforms in the RA.

Highly Scalable Platform for the Construction of Dependable and Ubiquitous Services: Spanish Ministry of Education and Science (TIN2007-67353-C02). Analysis and tests of consistency problems that arose in end-user applications when second-level caches (e.g. Coherence, JBoss cache) are combined with object persistence mechanisms (e.g. Hibernate).

AUTOMAN - Autonomic Management of Grid-Based Enterprise Services: Did the integration of self-configuration and self-repair properties of autonomic computing in the core of a cloud platform at INRIA (France).

High Performance Distributed Systems: Community of Madrid (S-0505/TIC/000285). Developed a high-available and scalable service for the JOnAS J(2)EE application server. It provided high availability for critical applications deployed in application server clusters, scaling-out the cluster when overloaded. UPM & BULL signed a pre-agreement to include it in the commercial version. LoC Java (HA&S Service) ≈ 4,000.

S4ALL (Services for All): 5th European Framework Programme (IST-2001-37126). Developed a high-available service for the JOnAS application server maintained by Bull SAS (France). Available since v.4.8. LoC Java (JOnAS) ≈ 150,000.

AUTONOMIC: Autonomic, Dependable and Middleware for Scalable, Distributed, Ubiquitous and Highly Available e-Services: Spanish Ministry of Education and Science (TIN2004-07474-C02-01). Built an open-source reference implementation of the WS-CAF specification for adding transactions to SOAP Web Services. LoC ≈ 50,000.

ADAPT: Middleware for Adaptive and Composable Distributed Components: EUREKA/ITEA project (Label 04025). Implemented a transactional-aware replication architecture for stateful EJBs for JBoss. LoC Java ≈ 10,600. Open-sourced the implementation of the Activity Service specification to add advanced transactions models to J2EE. LoC Java ≈ 10,000.