Engineered memory-optimized data pipelines to process 30,000 raw hardware oscilloscope recordings. Diagnosed and resolved a 29GB CPU RAM crash by squashing batch and window dimensions. Isolated a leaking signal at AUC ~0.54 across 5 neural network architectures.

Solved a non-convex MIQCP optimization problem over a 288-node, 287-arc real gas network in Pyomo and Gurobi. Tightened flow boundaries to ±50 kg/s to eliminate solver timeouts and generated 50+ multi-objective Pareto fronts.

Engineered an Agentic RAG architecture using LangGraph for autonomous intent routing, combining Dense embeddings (ChromaDB) and Sparse matching (BM25) with Reciprocal Rank Fusion.

Designed strict LLM-as-a-judge reflection loops to actively prevent hallucinations.

Deployed to Google Cloud Run by caching 1.5GB HuggingFace models directly into Docker image layers, successfully reducing container cold starts from 105s to <5s.

View Live Prototype →

Developed a quantile LightGBM forecasting system that extracted 17,000 hours of market variables into PostgreSQL, outperforming a naive persistence baseline by 35.6% MAE (10.82 €/MWh).

Engineered an interactive Tableau dashboard to synthesize massive datasets into clear, strategic narratives.

Shipped the full-stack architecture to production on Google Cloud Run via automated GitHub Actions CI/CD workflows.

View Live Prototype →

Built an AI agent in 24 hours that evaluates optimal European data center locations. It cross-references plain English requirements with live data from Ember, IEA, PyPSA-Eur, and OpenStreetMap to solve complex constraints around cost, clean energy, and grid capacity.

View Live Prototype →

Achieved mAP of 0.78 on ICDAR17 dataset. Boosted baseline by 24% via Generalized Max Pooling (GMP) and PCA-whitened Multi-VLAD ensemble.

Combined multi-scale Selective Search, ResNet18 CNN, and Linear SVM to reach MABO of 0.74.

Extracted 3D dimensions from noisy ToF sensor data using MLESAC plane-fitting.