I build systems that can think, listen, and see.
Focused on creating practical AI applications and tools for analyzing multimodal content.
Extracting insights from video content with local LLMs and developing interactive systems that understand visual and textual data.
BloomFilter, TDigest, HyperLogLog, CuckooFilter, and CountMinSketch are my favorite Algebras.
Kaggler. https://www.kaggle.com/vincento. @HackerSchool alumnus. Spaces+VIM.
Detecting when AI systems strategically comply while preserving hidden values. Fine-tunes GPT-OSS 120B using QLoRA to identify self-preservation reasoning patterns in model decision-making. Addresses the critical AI safety concern from Anthropic's research on models that behave as aligned during training while harboring different values. Includes pre-trained af-detector-120b-v3 model and comprehensive evaluation tooling.
Comprehensive AI safety project implementing GuardReasoner, a state-of-the-art reasoning-based LLM safety classifier achieving 84% F1 score. Features two-stage training (R-SFT + HS-DPO), cost-effective data generation using Gemini 2.0 (600× cheaper than GPT-4o), and transparent step-by-step reasoning for safety decisions. Currently achieving 59% accuracy with LLaMA 3.2-3B, targeting 80-85% with full pipeline.
Exploring diffusion models for text generation, applying techniques from image generation to natural language. This work investigates novel approaches to controlled text synthesis using diffusion processes, bridging computer vision and NLP methodologies.
Cataloging and analyzing adversarial patterns, prompt injection techniques, and security vulnerabilities in large language models. This research focuses on identifying, documenting, and mitigating abuse vectors to build more robust and secure AI systems.
A modern Python/Flask event processing system for compliance monitoring, combining probabilistic data structures (HyperLogLog, Bloom Filters, TopK, Count-Min Sketch) with Algebird-style monoid abstractions. Enables memory-efficient real-time analytics for security and audit logging with privacy-preserving compliance for GDPR/CCPA scenarios.
Combining my security background with AI capabilities, I've developed tools for monitoring AWS CloudTrail logs, anomaly detection for compromised logins, and ML-based security monitoring solutions. Recently, I've created an MLLM-based Incident Response Agent to speed up evidence collection after data breaches.
Developing tools to monitor and analyze LLM responses and interactions. The chatMonitor project helps review and evaluate responses from language models, providing insights into their performance and behavior when processing video content.
My recent work focuses on multimodal AI, LLMs, and developing robust evaluation frameworks for testing language models. I've created several benchmarks and evaluation tools to assess LLM performance across different domains. I primarily code in Python, with some Scala and C++ for specific applications. Here are some of my highlighted projects:
Detecting deceptive AI behavior using GPT-OSS 120B + QLoRA. Identifies self-preservation reasoning in LLMs that comply strategically while preserving hidden values.
Learn moreState-of-the-art reasoning-based LLM safety classifier implementing GuardReasoner. Two-stage training with cost-effective Gemini 2.0 data generation (600× cheaper than GPT-4o).
Learn moreDiffusion models for text generation, applying image generation techniques to natural language processing.
Learn moreComprehensive catalog of adversarial patterns, prompt injection techniques, and security vulnerabilities in LLMs.
Learn moreEvent processing system using probabilistic data structures for compliance monitoring with privacy-preserving analytics.
Learn moreMonitoring tool for videoLLM interactions that analyzes chat patterns and performance metrics.
Learn moreCreating synthetic data for training specialized language models with focus on domain-specific knowledge.
Learn moreMLLM-based IR agent that speeds up evidence collection after data breaches and provides automated security assistance.
Learn moreComprehensive evaluation framework for assessing LLM capabilities, biases, and limitations with standardized test suites.
Learn moreEvaluation framework for benchmarking mathematical reasoning capabilities of language models and neural networks.
Learn morePython; PyTorch; Diffusion Policies for robots; Vision LLMs; LLM Evaluation Frameworks; Computer Vision; OpenCV; AWS EC2, S3, Lambda; Azure Data Warehouse
I'm a gun slinger turned code slinger, writing stuff that humans & computers can read. Worked at RCMP, CIBC, and Deloitte. @TechStars Chicago 2013. Proud @HackerSchool alumnus Winter 2012. Competition Expert level Kaggler. https://www.kaggle.com/vincento
Focused on building state-of-the-art machine perception solutions for robotic vehicles and interactive AI systems.
SecOps -> CISO -> Security Engineer -> Data Engineer -> Machine Learning Engineer
Find me on Github