Allan Sevilla

BSA/AML Analyst & AI Automation Developer @ Santander Private Banking  ·  Miami, FL
BS Computer Science, FIU (Cum Laude)  ·  Building at the intersection of applied AI and financial-crime compliance

 

I'm a Computer Science graduate from FIU (Honors College, Cum Laude) working as a BSA/AML Analyst and the sole automation developer for the Transaction Monitoring Unit at Santander Private Banking, where I build LLM-driven investigation and alert-triage pipelines for financial crime detection.

My work spans agentic AI pipelines, federated learning for fraud detection, deep learning classifiers, automated ML workflows, and OS-level concurrency in C. I enjoy building things that require both rigorous engineering fundamentals and cutting-edge AI — from Banker's Algorithm implementations to multi-agent, graph-based retrieval and investigation systems.

Languages   Python · C · SQL · VBA · Bash · JavaScript · HTML/CSS

ML / AI   LangGraph · LangChain · Anthropic API · Pydantic v2 · TensorFlow · PyTorch · scikit-learn · OpenAI Gym · RAGAS · Ollama

Databases & Infra   Neo4j · ChromaDB · PostgreSQL · Redis · Docker · AWS Lambda · Streamlit · FastAPI

Tools & Environment   Linux · Neovim · Hyprland · Git · Google Colab · uv

Agentic alert-triage system combining deterministic risk-tiering logic with LLM-generated case narratives (LangGraph, Pydantic v2, FastAPI, Neo4j, Anthropic API), validated against synthetic data with planted AML typologies like structuring and funnel accounts.

Full-stack AI-powered task management app with voice input. Built with a React frontend, FastAPI backend, PostgreSQL database, and local LLM integration via Ollama (Llama3). Users speak tasks naturally — the AI parses and structures them automatically.

End-to-end automated ML workflow with auto task detection (regression vs. classification), data auditing, robust preprocessing, outlier handling, feature selection, a multi-model zoo with hyperparameter tuning, and optional ensemble blending — all through a single configurable class.

Multi-agent financial analysis system built around a LangGraph-style architecture with composable agents, graphs, prompts, and services for graph-driven AI reasoning over financial data.

Undergraduate research project (ARCH Honors Program) on Bitcoin fraud detection using multi-party vertical federated learning with PU learning. Achieved 0.974 ROC AUC, matching a centralized oracle baseline while preserving institutional data privacy.

Personal portfolio — allansevilla05.github.io

• SantanderAI/gen-fraud-graph — Authored and merged a StructuringGenerator class implementing BSA structuring typologies into a bank-maintained synthetic fraud-graph generator (PR #17), with full test coverage and GPG-signed commits.

• Financial Crime & Compliance Automation — LLM-based alert triage, case narrative generation, and transaction monitoring workflows in production
• Artificial Intelligence — reinforcement learning (TD Learning, Q-Learning, SARSA), custom OpenAI Gym environments
• Deep Learning — CNNs with TensorFlow, image classification for waste sorting
• Operating Systems — concurrency, resource management, Banker's Algorithm in C, priority scheduling
• Research — federated learning, graph-based RAG systems (LangGraph + Neo4j + ChromaDB)