Dilrabo

Independent AI Research & Applied Machine Learning

Independent Reserve • Trabajador autónomo

Dec 2025 - Present • 5 mos

Wildfire Risk Prediction using Uncertainty-Aware Spatiotemporal GNN PyTorch Geometric · Graph Learning · Geospatial ML · Uncertainty Modeling • Designed and deployed a 327K-node graph-based wildfire prediction system from multi-source raster data using 8-neighborhood spatial connectivity, enabling large-scale geospatial reasoning. • Built an end-to-end ML pipeline including raster alignment, feature engineering (61 features), graph construction, and model training, ensuring scientifically valid spatial relationships. • Implemented and compared GAT, GCN, and GraphSAGE architectures, achieving up to R² ≈ 0.76, outperforming strong baselines (XGBoost, Random Forest, CNN). • Addressed stochastic simulation label noise by integrating Gaussian NLL loss, modeling aleatoric uncertainty directly a key limitation in traditional wildfire prediction systems. • Developed uncertainty estimation using MC Dropout (30+ stochastic passes), enabling per-node confidence estimation for safety-critical decision-making. • Performed calibration analysis and temperature scaling, reducing miscalibration and aligning predicted confidence with actual coverage (reliability curves, PICP). • Designed a counterfactual intervention framework (fuel reduction, firebreak simulation), revealing critical limitations in feature design and enabling future causal wildfire modeling. • Identified and resolved data engineering challenges including raster misalignment, nodata inconsistencies, and spatial leakage, ensuring robust and reproducible modeling. AI Voice Agent for Appointment Automation | STT · TTS · FastAPI · Telephony Integration (Process) • Developing an end-to-end AI voice agent for appointment booking and customer interaction automation. • Implementing speech-to-text, intent detection, and dialogue management pipelines for real-time conversational workflows. • Integrating voice synthesis and backend scheduling tools to simulate AI-powered receptionist functionality. • Designing the system

ML Engineering Intern Production AI Systems

IndustryPath • Tiempo completo

Aug 2025 - Oct 2025 • 2 mos

• Architected and shipped BlazeVeritas AI, a production-grade LLM + computer vision system FastAPI backend, Streamlit frontend, Azure Cloud deployment serving real-time wildfire intelligence at >95% detection reliability. • Engineered a hybrid RAG pipeline (BM25 + ChromaDB) with custom retrieval benchmarks; built FastAPI endpoints enabling sub-second inference serving. • Optimized CNN inference pipeline achieving ~40% speed improvement through batch processing, model quantization awareness, and efficient data loading without accuracy regression. • Implemented production observability: MLflow experiment tracking, MC-Dropout uncertainty flags, Grad-CAM explainability overlays, and automated CI/CD via GitHub Actions. • Maintained reproducible ML environments versioned data pipelines, and structured model registries enabling team-wide replication of all experiments.

ML Engineering Intern LLM Pipelines & Automation

Start-Up Business Consulting • Tiempo parcial

Nov 2024 - Feb 2025 • 3 mos

• Built and deployed an LLM-powered conversational AI assistant reducing manual dispatcher/recruiter communications by ~40%; integrated into live operational workflows with N8N automation. • Engineered ML pipelines for ETA prediction and logistics delay modelling using XGBoost and feature stores, improving on-time delivery prediction accuracy by ~12%. • Designed and ran structured prompt evaluation cycles in Azure AI Studio, iterating on system messages and few-shot templates to measurably improve agent output quality and stability. • Integrated Parler-TTS and OuteTTS voice AI models into backend automation pipelines, extending system functionality without re-architecting core services.