I build and optimize high-performance AI systems, from large-scale data pipelines to generative models. I'm passionate about turning complex data into actionable insights and reliable, production-ready software.
- High-Throughput Data Systems: Engineering distributed query and data pipelines on multi-billion-row warehouses (Trino, SQL).
- Retrieval-Augmented Generation (RAG): Implementing and optimizing RAG pipelines to enhance LLM performance and reduce operational costs (LlamaIndex, Pinecone).
- Generative AI Models: Building and deploying Generative Adversarial Networks (GANs) for high-fidelity synthetic data generation (TensorFlow).
- Real-Time Predictive Analytics: Developing and validating models (RandomForest, Scikit-learn) for time-critical decision-making in high-stakes environments.
- End-to-End MLOps: Containerizing services (Docker) and building full CI/CD pipelines (GitHub Actions) for automated testing and deployment.
| Company / Project | Key Contribution & Impact | Technologies Used |
|---|---|---|
| Viasat ML Engineer Intern |
Engineered a distributed Text-to-SQL pipeline, reducing query latency by 2.5x and cutting LLM costs by 42% using RAG. | Python, Trino, SQL, RAG, Docker, CI/CD |
| FY Motorsports Data Science Intern |
Built real-time telemetry pipelines and a RandomForest model (80%+ accuracy) to optimize professional race strategy. | Python, Scikit-learn, Pandas |
| UCLA AI Hackathon 1st Place Winner |
Led a team to engineer a GAN that achieved 99.87% distributional similarity for a 7M-record private dataset. | TensorFlow, Python, Docker, Intel TDX |
| Esports Analytics Engine Project Lead |
Developed an LLM-powered recommendation engine on AWS to generate optimal esports team compositions. | AWS Bedrock, Python, S3, REST API |
- Languages:
Python,C/C++,SQL (PostgreSQL),R - ML/Data:
PyTorch,TensorFlow,Scikit-learn,Pandas,NumPy,LlamaIndex,RAG,Trino - Cloud/Tools:
AWS (Bedrock, S3),GCP (BigQuery),Docker,CI/CD,Git,Flask,React
- LinkedIn: linkedin.com/in/lucaschin1
- Email: lucaschin0405@outlook.com