🛠️ Staff AI Developer @ Tiptree Advanced Systems | 🎓 Ph.D. in Computer Science, McGill University
🤖 AI Agent Systems · 🛡️ AI Safety & LLM Security · 🧠 LLM Training & Post-training

I build and secure AI agent systems. At Tiptree Advanced Systems I work on the infrastructure that research agents run on: sandboxed code execution, job orchestration across HPC clusters, the credential and access model agents operate under, and deep-research agents with tool access. Before that, I was a postdoctoral researcher at McGill working on LLM security, agent safety evaluation, and data-efficient LLM training.

My research record spans AI agent safety and LLM security (constraint-violation benchmarks, prompt injection, jailbreaks, red teaming), LLM pre-training and post-training, and, earlier, adversarial robustness, interpretability, and malware detection. I care about the same question from both sides: how to make agents capable, and how to know when they go wrong.

• 🤝 Althea
  An always-on AI research assistant with a persistent agent runtime. I build the runtime it runs on (sandboxed code execution, job orchestration across HPC clusters, and the credential and access model agents operate under) and its deep-research agent.

• 🗺️ Lacuna
  A research map for machine learning that surfaces under-explored research directions from the literature: directions, papers, authors, and venues. I built its hybrid lexical + semantic search, the agent-facing context APIs, author profile claiming with verified identity, the new-papers feed, the hosted MCP endpoint, and the production database and search-index operations. Here is my researcher profile on Lacuna.

• 🔌 Lacuna Research MCP
  Open-source MCP server that gives coding agents and assistants access to Lacuna: search research directions, papers, authors, and venues, and generate research proposals.

• 🖥️ tasc-hpc-daemon
  PyPI package: a daemon that bridges AI agents to remote HPC compute, so an agent can submit, monitor, and collect jobs on a cluster it does not run on.

• 🧪 ODCV-Bench
  A benchmark for evaluating outcome-driven constraint violations in autonomous AI agents: whether agents break explicit rules when doing so helps them reach their goal. The paper reached #1 on Hacker News in February 2026 (544 points, 366 comments; ranking history).

• 🛠️ Kam1n0-Community
  Scalable binary similarity analysis platform for reverse engineering, built with Apache Spark and Cassandra.

• 🧮 iffnn
  Interpretable feedforward neural network architecture for tabular data with intrinsic explanations for its predictions.

• 🦙 DMaS-LLaMa-Lite-Training-Code
  Complete training pipeline for a 1.7B-parameter LLaMa-based model, optimized for data-efficiency.

• 🔁 Incremental-LLM-Training
  Layer-wise incremental LLM training: a novel paradigm to improve compute utilization and training throughput.

COMPASS-Implementation (multi-agent system after Google Research's COMPASS, with MCP tools and a FastAPI dashboard) · Miles_AI_Collection (LLM training and LangChain experiments) · Virtual Customer Service Representative (LangGraph ReAct + FAISS RAG app) · Miles_Data_Science_Collection (EDA and modeling on real-world datasets)