Felix – Python, OpenAI, LangChain

It's an internal AI search and knowledge retrieval platform for a large enterprise document base. The system enables teams to search across millions of PDF documents using semantic retrieval, LLM-based answer generation, and observability tools for quality tracking.

• Built a RAG system for internal search over 2M+ PDF documents.
• Improved retrieval quality with semantic chunking, Pinecone vector storage, and RAGAS evaluation.
• Designed FastAPI back-end services for async LLM requests, retrieval workflows, and tool-calling logic.
• Used Celery for non-blocking long-running LLM tasks and better system responsiveness.
• Exposed back-end endpoints to a TypeScript frontend for internal business teams.
• Integrated Langfuse to track prompts, completions, latency, and retrieval quality.
• Tuned embedding retrieval and Elasticsearch access patterns to reduce read latency.

It's a computer vision solution for automated visual defect detection using custom-trained object detection models. The system replaces manual visual checks and improves consistency, scalability, and detection performance.

• Trained and deployed custom YOLOv8 models on a 500GB image dataset for visual defect detection.
• Improved F1-score from 0.72 manual baseline to 0.91 through optimized model training.
• Built OpenCV and Python preprocessing pipelines to clean and prepare image datasets.
• Migrated large-scale image data to Azure Blob Storage with Docker-based model deployment.
• Collaborated with data and operations teams to replace manual inspection workflows.
• Monitored model quality and supported retraining based on production feedback and drift.

It's a large-scale model monitoring and data quality pipeline that tracks feature distribution changes, detects model drift, and triggers retraining workflows for production machine learning systems.

• Built Apache Spark jobs in Databricks to monitor model drift across 50M+ daily records.
• Created drift detection rules to trigger retraining when feature distributions shifted significantly.
• Built data pipelines for research and production datasets used in model training and monitoring.
• Orchestrated recurring data processing and quality validation workflows with Airflow.
• Managed infrastructure components with Terraform.
• Added observability with OpenTelemetry and New Relic for production AI/ML workflows.
• Improved ML system reliability by catching data quality issues before they affected predictions.

It's production machine learning pipelines for churn prediction and business analytics. The project involves processing high-volume transaction data, preparing training datasets, deploying prediction APIs, and evaluating models in production traffic.

• Processed 5M+ daily transaction records using Snowflake and AWS for churn prediction.
• Cleaned and transformed data with Python/Pandas and built SQL analytics across customer segments.
• Containerized ML models and APIs with Docker, Kubernetes, and KServe for production serving.
• Optimized vLLM inference to cut deployment time 50% while reducing costs.
• Ran A/B tests comparing new model versions against production baselines.

It's automation tools for image tagging and preprocessing to reduce manual labeling effort and accelerate the preparation of computer vision datasets.

• Automated image tagging and preprocessing with OpenCV Python scripts.
• Reduced manual tagging time from 40 hours/week to 5 hours/week.
• Built reusable data processing workflows for ML model preparation.
• Standardized image metadata and labeling for consistent training.