Josue Valenzuela Perez

1st Place — Hola Mundo 2026 National Innovation Competition. Production-ready AI video analytics platform supporting 12 concurrent streams with 872ms E2E latency. Multi-model pipeline: YOLOv8 object detection, pose estimation (fall detection), and custom fine-tuned weapon detector (0.770 mAP@50). Curated 14K-image weapon dataset.

Fine-tuned Vision Transformer on FaceForensics++ (7K videos, 6 manipulations). Achieved 0.972 test AUC, competitive with Xception (0.978), outperforming EfficientNet-B4 (0.967). Demonstrated superior robustness under heavy blur (AUC 0.873 @ K=7 vs 0.821 / 0.781) and stable JPEG performance. Cross-dataset eval on Celeb-DF v2.

ROC Curves

AUC per Method

Robustness

Fully local RAG system indexing 1,557 pages of medical literature into 8,210 embeddings. ChromaDB vector search with semantic chunking achieves <250ms retrieval latency and L2 distance of 0.45. Offline quantized LLM inference eliminates cloud API costs entirely.

$ python ask.py "What precautions for mobilizing a patient with drains and catheters?"

Based on the surgical nursing reference:
• Secure all drains and catheters before moving
• Keep drainage bags below insertion site
• Avoid tension on tubing during transfer
• Monitor for pain, leakage, or dislodgement

Semantic segmentation of aerial imagery (Potsdam dataset, 5cm/px) using a CNN and an encoder–decoder architecture with skip connections. The encoder–decoder significantly outperformed the baseline CNN (77.2% vs 63.4% test accuracy) across 6 land-cover classes.

RGB

Elevation

Ground Truth

Prediction

Implementation of Gatys et al. (2015) with three extras: progress GIF saving, two-style blending with controllable weight, and color preservation mode. Supports non-square outputs with aspect-ratio preservation and a full CLI interface.

Content

Style

Result