[{"content":"This poster was presented at the 2025 Silicon Valley Cybersecurity Conference (SVCC). The project develops neural architectures for producing person-specific 256-bit cryptographic keys from ECG segments. Evaluation focuses on consistency across recordings, separation between individuals, and randomness under the NIST Statistical Test Suite.\nThe published architecture combines convolutional feature extraction with multihead attention.\n","permalink":"https://luciano665.github.io/publications/ecg-key-generation/","summary":"Learning person-specific 256-bit keys from ECG signals with attention to consistency, separation, and statistical randomness.","title":"A Deep-Learning Approach for ECG-Based Cryptographic Key Generation"},{"content":"Agentic Support routes customer-service requests across FAQs, refunds, and troubleshooting. It uses semantic retrieval through Chroma, a modular FastAPI backend, and Langfuse traces to inspect agent decisions and tool calls.\n","permalink":"https://luciano665.github.io/projects/agentic-support/","summary":"A support workflow combining retrieval, tool use, and a ReAct agent with inspectable execution traces.","title":"Agentic Support"},{"content":"PrivDrift studies privacy risk inside an active model context. It evaluates whether a secret disclosed by a user can be elicited later after unrelated, content-dense turns and persuasion-based probing.\nI led the research design and developed the hybrid evaluation pipeline for detecting direct, partial, and paraphrased leakage. The manuscript evaluates 1,000 controlled dialogues per model across three language models. Within the tested drift window, additional topic drift does not reliably reduce leakage.\n","permalink":"https://luciano665.github.io/publications/privdrift/","summary":"A benchmark for testing whether user-disclosed secrets remain recoverable after a conversation moves to unrelated topics.","title":"PrivDrift: Auditing User-Secret Leakage Under Topic Drift in Active LLM Conversations"},{"content":"The system clones and parses repositories with Tree-sitter, produces syntax-aware chunks, and connects natural-language questions to relevant source code through vector retrieval.\n","permalink":"https://luciano665.github.io/projects/codebase-rag/","summary":"Natural-language exploration of code repositories using syntax-aware parsing and vector retrieval.","title":"Codebase RAG"},{"content":"This work compares distilled and non-distilled language models on reliability and sycophancy. It also studies whether conformal risk control can turn calibrated uncertainty thresholds into a practical intervention for harmful agreement.\n","permalink":"https://luciano665.github.io/publications/conformal-sycophancy/","summary":"Studying harmful agreement in language models and uncertainty thresholds designed to reduce sycophantic responses.","title":"LLM Sycophancy and Conformal Risk Control"},{"content":"This application organizes prompts, reference answers, and model outputs into a repeatable evaluation workflow. It combines scoring metrics with an LLM-based judge for cases that require semantic assessment.\n","permalink":"https://luciano665.github.io/projects/llm-evaluation/","summary":"A structured application for comparing model responses with similarity metrics and an LLM judge.","title":"LLM Evaluation"},{"content":"This project asks whether a medical vision-language model that agrees with a false spatial claim failed to perceive the image or chose to follow the user\u0026rsquo;s premise despite internally representing the correct answer.\nThe current benchmark contains 3,244 spatial questions derived from SLAKE annotations. Neutral prompts establish baseline competence before evaluation under six pressure conditions. For self-hosted models, linear probes test whether the correct spatial information remains recoverable from vision-tower representations.\nThe study spans six models. Human auditing and some generation issues remain under investigation, so the results are preliminary.\n","permalink":"https://luciano665.github.io/publications/medical-vlm-sycophancy/","summary":"Separating failures to perceive a medical image from agreement with a user\u0026rsquo;s spatially false premise.","title":"Sycophancy in Medical Vision-Language Models"},{"content":"PixelAlchemy provides an image-generation interface backed by models deployed on serverless GPUs with Modal. The application includes persistent storage and a responsive TypeScript frontend.\n","permalink":"https://luciano665.github.io/projects/pixelalchemy/","summary":"A text-to-image application connecting a web interface to generative models on serverless GPUs.","title":"PixelAlchemy"}]