← back

----------------------------------------| ****@@@@@@ 0 : 0 | *******@@@@@@@@@ | *********@@@@@@@@@@@@ : | ***********@@@@@@@@@@@@@@@ | ***********@@@@@@@@@@@@@@@@@@ # : | ***********@@@@@@@@@@@@@@@@@@@%%% # @ | ***********@@@@@@@@@@@@@@@@@@@%%%%% # : | ***********@@@@@@@@@@@@@@@@@@@%%%%% # | ***********@@@@@@@@@@@@@@@@@@@%%%%% +++++ # : | ***********@@@@@@@@@@@@@@@@@@@%%%%% ++++++++++++++++ | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ : | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ : # | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ # | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ : # | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ # | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ : # | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ : | ***********@@@@@@@@@@@@@@@@@@@%%%%%+++++++++++++++++++ | ***********@@@@@@@@@@@@@@@@@@@%%%%% . ++++++++++++++++ : | ***********@@@@@@@@@@@@@@@@@@@%%%%%. . . . . . . +++++ | ***********@@@@@@@@@@@@@@@@@@@%%%%% . . . . . . . . . : | ***********@@@@@@@@@@@@@@@@@@@%%%%%. . . . . . . . . . | ***********@@@@@@@@@@@@@@@@@@@%%% . . . . . . . . . . : | ***********@@@@@@@@@@@@@@@@@@. . . . . . . . . . . . . | ***********@@@@@@@@@@@@@@@. . . . . . . . . . . . . . : | .*********@@@@@@@@@@@@. . . . . . . . . . . . . . . . | . . *******@@@@@@@@@. . . ._. . . . . . . . . . . . . ----------------------------------------| . . . ****@@@@@@. . . . .|.|. . . . . . . . . . . . .

Pong and a DOOM corridor, two of the games the agents learn to play.

01 · Research group mentor · MSOE AI Club

World models

Training reinforcement learning agents that act by imagining the next states of their game, on Atari games and DOOM.

World models · Reinforcement learning · DreamerV3 · R2-Dreamer

My roleMentor of an MSOE AI Club research group.

The research question

A world model predicts the future internal states of the game it plays, a dream of what happens next. We study when a better dream stops making the agent better at winning.

Seeing the dream

For Pong we use the decoder-based DreamerV3 reproduction in the R2-Dreamer codebase. Its decoder turns predicted states into images, so we can compare the dream with the real game frame by frame. Our open-loop videos reconstruct five observed frames, then predict what follows from the recorded actions.

What we have seen so far

The dream converges fast. In recent Pong runs the board appeared after about 30 steps, the paddles after about 100, and the ball at about 500. Up to 800 steps we have not yet seen a notable change.

. -. .. =@* .#. .# :: * .-=-: .@+. -*. . .: =- - :-. --.:.. .+.. .. .: . .+=. .+= *::. * .- .. : :+ :%*. :. :*.+ .- :. .= :- -. .##. . . =: -: :#- - .-. . =: .: :# . .# :: . .: : .+ - = .- .*.#. : . : : +: .. +. - *. * .. - =. .: . . ..::.. :: *: - :- -- = - .:-==-:. .: :. .. - .= .-+##+-. . * -. . .- .:=#%%#=:..: - -. .- .:=#%%#=:..- -= =. : -..-+##+-. .: .-:- =+ : . :..:-==-*. .+:+.: -. *# ..::.+- :+ . = -. .# :: = .: : . .* .*. ::* .: +: % .# * .@: .- =:. . .: == :* . -. : = =. : .. -.: .# *-- .. .. .#-- : -- * :#. . .. . .+ +. =+ -- + :* .+ - *: :-- .. =: :. - - .- .+* : .: -* .=: :%. .-: = .+ . :: := :- .- -

Stars drifting behind a galaxy cluster are bent into arcs by its gravity.

02 · Undergraduate research · Prof. Adam Wright

Galaxy cluster cosmology

Studying how blended galaxies affect cluster mass estimates, and how machine learning could detect them.

Weak gravitational lensing · Source blending · Machine learning

My roleUndergraduate researcher with Prof. Adam Wright at MSOE.

The research focus

In crowded cluster fields, overlapping galaxy images can affect shape measurements. The research looks at that blending and what it means for weak-lensing mass estimates.

Building the foundation

I'm learning the lensing concepts and working through the existing pipeline, focused on how the shape measurements are made.

Detecting blends with machine learning

Next, I'm looking to use machine learning to detect blended sources, so overlapping galaxies can be flagged before their shapes feed into mass estimates.

Why it interests me

It ties the physics of the universe to practical questions about images, data, and computation.

← back

+----------------------------------+-----------------------------+-----------------------------+ | | | | | ======== | | === | | ======== | | === | | -:::- | | | | - - | | | | - - | | | | : : |............... | | | : @...:.............. ...............................x | | : : . | | - - . | | - - | | . | | -:::- | | . | | | | . | | | | . | +--------------- ---------------+ +------------- . ------------+ | | . | | | . | | | . | | | . | | | . | | | . | | | .... | | | | | | | | ======== | === | | ======== | === | | | | +----------------------------------+-----------------------------------------------------------+

A robot patrols an ASCII floor plan.

03 · Undergraduate team research · Aug 2025 – May 2026

Digital twin & robotics

Training a robot on a building from its architectural model, before it ever drives the real halls.

NVIDIA Isaac Sim · OpenUSD · ROS 2 Nav2 · Vision-language models · Rosie supercomputer

My roleI built the twin's physics layer: colliders for floors, walls, glass, and stairs, and door joints with calibrated hinges and motor torque. Everything downstream runs on it.

Buildings already have a blueprint

Service robots usually train on a finished, live campus. Architects already model every wall and door in Revit, so our team turned the models of six MSOE buildings into simulation-ready twins in NVIDIA Isaac Sim.

A model has to behave

A door needs a working hinge, not the look of one. I added triangle-mesh colliders to the static structure, classified the door mechanisms, and gave every hinged door a revolute joint and drive so it swings open.

What the team built on top

We ray-traced occupancy maps of each floor and drove a door-inspection patrol with an NVIDIA Nova Carter robot on ROS 2 Nav2. A Qwen3-VL model fine-tuned with LoRA spots open doors and cracked glass, the two issues campus safety cares about most.

Trained only on synthetic images

In the twin we opened doors and cracked glass on purpose under randomized lighting, materials, and camera angles, with labels for free. Trained on zero real images, the model reached 93% accuracy on photos of a real campus building. The project ran more than 3,500 GPU jobs on MSOE's Rosie supercomputer.

Where we presented it

Our team presented the project at MSOE's Rosie Competition and at MICS, the Midwest Instruction and Computing Symposium.

← back

| loss (log) | run 14 step 220 loss 0.709 | | run 14 step 280 loss 0.668 |-_ | run 14 step 340 loss 0.459 | -¯-_ | run 14 step 400 loss 0.419 | :-_ | run 14 step 460 loss 0.306 | ¯ ¯---_ | run 14 step 520 loss 0.220 | -- : | run 14 step 580 loss 0.170 | _¯____ | run 14 step 640 loss 0.146 | -__- ¯ | run 14 step 700 loss 0.129 | __-¯- _ | run 14 step 760 loss 0.095 | -- ¯ _ | run 14 step 820 loss 0.071 | ¯¯--¯¯-_ | run 14 step 880 loss 0.064 | - _ | run 14 step 940 loss 0.053 | - ---_ | run 14 step 1000 loss 0.046 | -¯_ -_ | run 14 step 1060 loss 0.040 | -- - | run 14 done ckpt saved | ¯_-__ | run 15 start lr 1e-3 | ¯- | run 15 step 40 loss 1.955 | ¯-_ | run 15 step 100 loss 1.495 | _ __ | run 15 step 160 loss 1.234 | ¯ -_ | run 15 step 220 loss 0.920 | _- | run 15 step 280 loss 0.766 | ¯-¯_ _ | run 15 step 340 loss 0.570 | ¯ | run 15 step 400 loss 0.505 | | run 15 step 460 loss 0.402 | | run 15 step 520 loss 0.359 | | run 15 step 580 loss 0.254 |. . . . . . . . . . . . . . . . . . . . . . . . . . . | run 15 step 640 loss 0.214

A metric curve next to a scrolling agent log.

04 · Personal project · In development

Turing

A workspace for the way I build and run research with AI agents.

Research agents · Tauri desktop · Python · Experiment tooling

My roleI set the direction, specs, integration, and evaluation, and use coding agents throughout.

Keeping the work visible

Terminals, agent sessions, live experiment metrics, and results share one keyboard-first desktop workspace, so I can follow what agents do while a project runs.

A goal and a way to check it

The research-agent framework pairs a goal with a fixed verifier. Attempts are bounded, checkpointed, and inspectable, so progress has to pass a defined check.

What I am exploring

I'm building the orchestration and evaluation workflow and testing how agents can improve it, which needs controlled evaluation.

← back

*****+************+++++** **++++++++++++++++++++++++++* *+++++++++++++++++++++++++++++* ++++++++++++++++++******++++++++* ++++++** ++++++++ *+++++* ++++** *****++== +++++ ++++** ##****+== ++++* *+++** @@%%##*+-- *+++* ++++* %%@@@%#*+-- +++* ++++ %%@@@@@@*+:: =++* +=++* #%%@@@ +*.: +++* ++++* ##%%%@@ =+.: +++* ++++* ####%@@ :=.. *+++* ++++ **###%@ :=.: *++** ++=+ ****#%@ --.: *+++* ++++++***# -:.. *++++** +++++++*# --:.:.:-====+++++** ++++++*# ---.....::---==++* =+++++* =--:....:--===++* ++++++ =---=-===++++** ++++++* ==++ ++++++* =+++ +++++* ++++ ++++++** ##*++++ +++++++++++++++*++ *+++++++++++++*++ **+++++==+++++++ *+++++**#*+**

The NCA trained for this site, running in your browser. This view shows a hidden channel; it damages itself and heals.

05 · Personal learning & experimentation

Neural cellular automata

Learning how local update rules can produce growth, regeneration, and coordinated behavior.

Neural cellular automata · Python · Visual experiments

My roleA personal learning project, built with coding-agent help to work through the math and behavior step by step.

Start with a local rule

It builds from cellular automata to trainable update rules. Each cell sees only its neighbors, yet the grid learns to grow a recognizable shape.

Then disturb the system

The experiments test damage, how training changes regeneration, and what the hidden channels hold. Later extensions cover robustness, initialization, and pathfinding.

Look beyond the training loss

I care about the gap between a model optimizing its objective and behaving the way we want. The notebooks show intermediate states, so growth, failure, and recovery can be inspected.

Based on

Alexander Mordvintsev, Ettore Randazzo, Eyvind Niklasson, and Michael Levin, “Growing Neural Cellular Automata”, Distill, 2020. My experiments and the model on this site follow the method from this paper.

← back

+-----------------------------------------------------+ | ? what is sepsis? | +-----------------------------------------------------+ sources +---------------+ +---------------+ +---------------+ | -------- | | -------- | | -------- | | | | | | | | ---- --- ---- | | ------ -- --- | | ---- -- -- -- | | ---- ----- -- | | ------ ------ | | -- ------ --- | | ---- --- ---- | | ----- ---- -- | | --- --- | > ===== ======= | | ------- | | ---- ----- -- | | --- ----- --- | | -- ---- ----- | > ====== | + plain summary ----------------+ | | | ---- --- | | ---- ----- -- | | | | ----- -- ---- | | --- ---- ---- | | | | | | --- -- --- -- | | --- ---- | | ---- --- ---- | | sepsis is the body's extreme | | ---- ---- --- | | | | -- ------ --- | --> | response to an infection. | | ------ --- -- | | ----- ------- | | ----- -- | | it can harm organs, so it | | -- ----- ---- | > ====== ====== | | --- -- ------ | ==== | | | -- ------ --- | | -- ------ --- | | -------- | | | > ===== ==== == | | ------ ------ | | | +-------------------------------+ | --- ------ -- | | --- -- ------ | | ----- -- ---- | | | | ----- ---- -- | | --- --- ----- | | --- ----- --- | | | | -- -- --- | | ------ ------ | | ----- ------- | | ------ --- -- | | --- ------ -- | | -- --- ------ | | ------ --- -- | | ----- -- ---- | | ----- -- ---- | | ---- ----- -- | | ---- -------- | | -- --- -- | | ----- -- ---- | +---------------+ +---------------+ +---------------+

A question searches the sources, pulls matching passages, and writes a plain summary.

06 · Team project · January 2026 · 2nd place

Hack 4 Health

Helping make biomedical text easier to understand, with source material close at hand.

Retrieval-augmented generation · Qdrant · Python · Team hackathon

My roleI built the retrieval layer: ingestion, embeddings, indexing, retrieval, and the RAG builder.

The challenge

At the MSOE and Medical College of Wisconsin hackathon, our team worked on patient-friendly summaries of biomedical text that preserve meaning and check the output.

My part of the pipeline

In the Database component I indexed health terminology and plain-language examples and retrieved context for generation, with tests around the retrieval code.

A team result

We placed second. Teammates built the validation layers and integration on top of my retrieval foundation.

← back

plan > work > review > verify +------------| agent------+ | | | | | | | | +------------| agent------+ | | |> lead------------+ +-------> review-------------------------> verify-+ | | | | | | | | | +------------| agent------+ | | | | | | | | | | | | | | | | | | +------------| agent------+ | | | | | | | +----------------------------------------------------------------------------------------------+

Packets move along a small graph of nodes.

07 · Personal workflow & tooling

Owen's Agent System

A consistent way to work with coding agents, from the first instruction to the final check.

Python · Terminal interface · Agent workflows

My roleI designed the workflow around how I use coding agents and built the launcher, shared instructions, and validation process with agent help.

One workflow, across agents

Work runs through Frame, Work, Prove, and Hand off. Shared instructions carry the task between tools, and adapters respect each agent's native controls.

Start from the terminal

The launcher picks a workflow, agent, model, and workspace for development, research, operations, tutoring, or explicitly scoped unattended work.

Make the next step clear

Task contracts, ownership boundaries, verification evidence, and checkpoints keep long sessions accountable. I judge it by the work it actually completes.

← back

Each skill is checked, then linked into the tools it belongs to.

08 · Open source tooling

Agent Skills

One place to write down how I want agents to work, shared by every coding agent I use.

Python · Claude Code · Codex · Cursor · Gemini · Obsidian

My roleI wrote the skills from workflows I actually repeat and built the installer, catalog checks, and tests with agent help.

Five copies, no source of truth

My skills had drifted across five tools. Some existed as separate copies that disagreed, some were invisible to the agent that needed them, and two claimed the same name. In August 2026 I cleared out 74 of them and started over in one repository.

Write once, link everywhere

Every tool reads the same format, so the problem was location, not format. A manifest maps each skill to the tools it belongs in, and the installer links it there. Removing a skill from the manifest unlinks it everywhere.

Skills that earn their place

A skill comes from a workflow I have already done by hand: sending jobs to MSOE's Rosie cluster, running reproducible experiments, or a professor mode that explains coursework without writing it. Catalog checks and tests run in CI on every change.

Owen Pacetti

Research

World models

Cosmology

Digital twin

Projects

Turing

Neural CA

Hack 4 Health

Agent System

Agent Skills

Experience

Cognex Software Engineering Intern

Blog

About

Contact

The research question

Seeing the dream

What we have seen so far

Next

The research focus

Building the foundation

Detecting blends with machine learning

Why it interests me

Buildings already have a blueprint

A model has to behave

What the team built on top

Trained only on synthetic images

Where we presented it

Keeping the work visible

A goal and a way to check it

What I am exploring

Start with a local rule

Then disturb the system

Look beyond the training loss

Based on

The challenge

My part of the pipeline

A team result

One workflow, across agents

Start from the terminal

Make the next step clear

Five copies, no source of truth

Write once, link everywhere

Skills that earn their place