Keshav Moorthy

Deep Learning & LLM

· Spearheaded a Knowledge Graph Generation and Extension task for a QA semantic search engine, generating and testing 64,000 novel high-quality entitiesfrom large unstructured scientific data, leveraging Transformer models with self-attention.

· Systematically tested and optimized the generated graph and search predictions using cross-validation, Hits@k and Mean Reciprocal Rank (MRR). Achieved nearly 5% improvement in F1 scores from initial models in entity predictions.

· Employed an LLM-prompting pipeline with GPT-3 into the transformer model for the Entity Linking task with some marginal improvement in F1-scores.

DevOps

· Implemented continuous integration and deployment (CI/CD) using GitHub Actions by automating deployment and testing for the system.

· Utilized Docker containers in an AWS environment, and integrated deployment monitoring tools like TensorBoard, Prometheus and Grafanafor tracking metrics.

· Reduced the Knowledge Graph search space query response time by 20% using FAISS and Text Mining techniques including integrated hash space algorithms.

Topic: L1-Personalized Lexical Complexity Prediction for L2-German-language-learners

· Developed an automatic personalized and contextualized word-difficulty prediction systemfor German-language-learners, using only prior language- and life-experiencesof learners as features.

· Achieved 29% MAE improvement with a novel BERT-based ensemble model architecturecompared to traditional models, and 0.71 Spearman correlation across a trial of 20 participants.

DevOps

· Designed and implemented an ELK stack (Elasticsearch, Logstash, Kibana) for real-time system diagnostics, improving response times for critical issues in production systems, in a C#-based automotive error-handling system for machine tools.

Team and Collaboration

· Collaborated and led an innovation project team of 3 interns in the Atlas Copco Global Industrial Internship program, Stockholm ’22.