D
M.Sc.Comp.Sci.

Daniel Lofeodo

PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO /
PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO / PORTFOLIO /

Latent SpaceMail: Agentic AI Research Newsletter

A weekly AI newsletter that writes itself. Six AI agents pick a standout research paper, sift through the week's tech news, and email a personalized briefing to subscribers. It runs on its own, watches its own quality, and I've been using it every Monday since May.

RAG With Receipts: Benchmarking RAG Accuracy, Grounding, and Latency

An AI question-answering system that shows its sources. Ask about Old School RuneScape and every answer links to the exact passage it came from, so you can check it. I tested it on 55 questions: it found the right passage 96% of the time, and says "I don't know" instead of guessing.

Autonomous Hydro-Dam Control: Managing 1-in-10,000-Year Extreme Flood Events with RL

An AI that learns to run a hydroelectric dam like a real operator, so engineers can test whether the dam survives extreme floods. Tested on 9 simulated floods (up to 1-in-10,000-year), it never breached safety limits and released 83% less spilled water than the industry software it was compared against.

RuneScape Infinite Prayer Plugin

A plugin for the game RuneScape that gives players unlimited prayer points by exploiting a timing quirk. It found 40+ paying customers through YouTube videos and forum posts alone.

Adversarial RL for Fast Nash Equilibria

A team project on getting game-playing AI to learn faster. We trained one agent to attack another's weaknesses, and the defender reached an unbeatable strategy in about 45% fewer training steps. Built in a custom tactical shooter game, with two teammates.

Exposing Spatial Blind Spots in Vision-Language Models

A test showing that top AI models, including GPT-5.4 and Gemini 2.5 Pro, struggle to imagine a scene from another angle. When asked what's left or right from a different viewpoint, GPT-4o does worse than random guessing at the hardest angles.

LLM-Driven Deterministic Clinical Guideline Compiler

Turns clinical guideline PDFs into step-by-step decision workflows that a computer can follow, with every step traceable to a line in the source. AI is used only while building them, never when they run, so the results are repeatable and safe to audit.

n2 × n2 Sudoku Solver

A Sudoku solver that handles puzzles of any size, from the classic 9×9 to 16×16 and beyond. It tries a number, backs up when it hits a dead end, and keeps going until the puzzle is solved.

Real-time NHL Goal Probability Prediction

A tool that predicts the chance an NHL shot becomes a goal. Enter a game ID and it analyzes the game's shots live, using machine learning models trained on real NHL data.

ABOUT / TOOLBELT / ABOUT / TOOLBELT /
ABOUT / TOOLBELT / ABOUT / TOOLBELT /

Hello!

I build production AI systems: agentic pipelines, RAG, and the MLOps around them. M.Sc. in Machine Learning from MILA (Yoshua Bengio's lab, Canada's premier AI research institute), 3 years as a software engineer, B.Eng. in Mechanical Engineering. I have shipped a live multi-agent LLM pipeline (LangGraph, Pub/Sub, Cloud Run) and a from-scratch RAG system with a 55-question golden eval, reranker benchmarks, and citation grounding checks. I monitor what I ship with drift, cost, and security tests, and I report null results honestly. AWS Certified Machine Learning Engineer (MLA). My roots are in Reinforcement Learning: at Hydro-Québec I built safety-bounded RL agents that stress-test dams under extreme floods. Reward design and rigorous evaluation carry over to everything I build. Here are some of the tools I work with:
  • Python
  • C++
  • C#
  • SQL
  • JavaScript
  • LangGraph
  • Claude API
  • RAG
  • FAISS
  • pgvector
  • LangSmith
  • PyTorch
  • scikit-learn
  • stable-baselines3
  • gymnasium
  • XGBoost
  • Pandas
  • GCP
  • AWS
  • Docker
  • Cloud Run
  • Pub/Sub
  • Firestore
  • Databricks
  • Spark
  • Weights & Biases
  • CI/CD
  • FastAPI
  • Firebase
  • Three.js
CONTACT / CONTACT / CONTACT / CONTACT / CONTACT / CONTACT /
CONTACT / CONTACT / CONTACT / CONTACT / CONTACT / CONTACT /