Lorenzo – Python, React, Typescript

Developed a high-performance encoder model. Introduced two key innovations for cross-lingual adaptation: (i) fusion of premium corpora and optimized tokenizers to minimize fertility issues, and (ii) intelligent weight initialization for accelerated convergence.

• Architected end-to-end training pipelines using PyTorch and Hugging Face Transformers.
• Integrated Weights & Biases (W&B) for comprehensive experiment tracking and reproducibility.
• Implemented Distributed Data Parallel (DDP) via HF Accelerate to enable efficient multi-GPU fine-tuning.
• Provisioned and managed NVIDIA GPU clusters for scalable training.
• Experimented with advanced NLP techniques, rigorously analyzed benchmarks, and delivered SOTA results on Brazilian Portuguese evaluations.
• Built the project entirely from scratch as the sole contributor.

Established an Event Driven Architecture from scratch, created an event analytics and replay system with S3 + Athena, built idempotent entrypoints—all sharply increasing system resilience. Implemented event-driven system ensuring no data loss even if IoT devices lost internet/power. This IoT project dramatically increased resiliency via async microservices, reducing data loss to essentially zero.

• Created scalable serverless APIs with Lambda (scheduled provisioned concurrency) and dashboard monitoring 1000s of functions; optimized cold starts (-70%) and costs (-40%).
• Built StepFunctions pipelines ingesting 10s of GBs in seconds for quick user access.
• Made entrypoints idempotent for at-least-once message brokers; used EventBridge choreography + StepFunctions orchestration for ML/image pipelines.
• Created/maintained Python SDK with async fire-and-forget for non-technical researchers, sharply speeding cloud feature delivery.

Co-founded an early GenAI startup and built pioneering RAG pipelines amid the rise of generative AI post-ChatGPT. Developed an enterprise chatbot for document interaction, experimenting with reranking, hybrid search, and retrieval-augmented generation (RAG) to enable natural conversations with company docs.

• Experimented and optimized RAG techniques, text chunking strategies, and data retrieval quality.
• Built custom loaders for PowerPoint, PDF, and other formats; implemented cloud ingestion pipelines.
• Engineered token streaming via Server-Sent Events and full pipelines with LangChain.
• Developed React/TypeScript frontend and constructed the entire platform from scratch.

Built from scratch to help users save on prescriptions: a price comparison engine for pharmacies/drugstores. Solo-developed with Flask (low-latency APIs), React, Jinja, PostgreSQL, Elasticsearch (ultra-fast search), JWT auth, and Scrapy scraping. Attracted 10k+ daily visitors and featured in local news.

• Designed PostgreSQL data model and med-matching heuristics across sites.
• Set up cron jobs for large-scale scraping + monitoring.
• Engineered Elasticsearch for extreme search speed.
• Delivered the complete platform end-to-end independently.