Lucas Dionisopoulos

Hey - I'm Lucas and I'm from Nebraska.

Here's my quick gist:

Initially studied finance (WashU) and worked as an investment banking for 2 years.
Knew I wanted a career change... so I took a gap year where I mostly studied (but also had a bit of fun).
Worked on research in grad school (UCSD) where I focused on RL for LLMs with PEARLS Lab. Published work I led at ICML (and gave a talk at a NeurIPS workshop on reasoning).
Currently: Jack of all trades at Hadrian (Special Projects).

If you're in LA / just want to chat - you should reach out!

Blogs

Some Writing

Why Pivot? My motivation for leaving finance, taking a gap year to study ML, and starting grad school.
My Hours as an Investment Banking Analyst. I tracked hours during my 2 years in banking and made some neat charts.
From Attention is All You Need to Llama 3. Technical overview of improvements to the decoder-only transformer used in Llama 3.
Nvidia's Multi-Trillion Dollar Moat: Explained. A technically approachable answer to this big question.

Research

Some Research

Primarily worked on RL + LLMs in PEARLS Lab. Previously worked with UCSD ProgSys.

How Reasoning Evolves from Post-Training Data: An Empirical Study Using Chess NeurIPS 2025 FoRLM Workshop (Oral) ICML 2026

Lead author. Outperformed state-of-the-art open-source reasoning models in chess through SFT and RL on a 7B-parameter language model. The key focus of this work was to study how fine-tuning influenced post-RL reasoning using custom theoretically-inspired datasets.

A* Neural Guided Search on ARC-AGI

Trained an image embedding model from scratch using a synthetic data generation pipeline to combine with a probabilistic context-free grammar in a neurosymbolic programming approach for ARC-AGI.

Philosophize That

Created a corpus of philosophical texts using various techniques to isolate signal from correlated noise and trained a model using document-level embeddings to classify the philosophical basis of my own journal entries.

Projects

Some Technical Projects

RL Benchmarking Gym

Built a clean, hackable RL gym engine from scratch to implement and ablate algorithms from baseline to SoTA.

Full-stack webapp combining image generation and image-to-3D asset models to create custom assets you can place in interactive dungeons or 3D print. Includes a dungeon generator that turns a prompt into a 3D dungeon using classical algorithms plus LLM agents. Built with React, Three.js, and self-hosted CV models.

SLAM-Based Autonomy for 4-Wheel Rover

Developed an autonomy stack for an uncalibrated 4-wheel differential-drive rover, implementing vision, EKF-SLAM from scratch, and path planning algorithms. Built with ROS2, OpenCV, and Python.

Reinforcement Learning Driver

Created and optimized a racing game in Pygame and taught an RL agent to drive it using Double Q-learning.

Machine Learning Anomaly Detection

Developed a real-time, interpretable anomaly detection system over noisy time-series data using kernel ridge regression for a San Diego healthtech company.

Recreating Word2Vec

Recreated Word2Vec vector arithmetic entirely from scratch, from building a tokenizer to cleaning English Wikipedia and training an embedding model.

Personal

Sabbatical

When I finished banking, I wanted to lock in and study ML.

I considered my options: study at home in Wisconsin or study abroad at long-term Airbnbs and pet sits.

One of these paths seemed more fun.

Note: I also did a lot of work.

Neat pics from a JMT attempt that was thwarted by wildfires.

Clean Our Green

Started a project in college to clean up a ton of St. Louis parks.

Contact

Feel free to reach out!

firstlast[at]gmail[dot]com