My lab notebook
RESEARCH
This is where I keep the work that began as a question. I am less interested in pretending every experiment was a breakthrough than in recording what I tried, what survived contact with reality, and what I would do differently next time.
PrototypeLatent world models for manipulation
Xlab-Jepa
Python, V-JEPA/V-JEPA2, SO-101 robotics
What it is
A robotics research scaffold for testing whether latent video representations can support action-conditioned prediction and planning for manipulation tasks.
What I wanted to know
Could a robot plan from a compact internal state without wasting effort rebuilding every pixel?
What I tested
Robot video ingestion, saved representations, action-conditioned prediction, and paths toward the SO-101 arm.
What I learned
The model was only part of it. The data, action labels, and evaluation setup mattered just as much.
PrototypeActive inference for robot control
taif
JAX, active inference, LeRobot, smolVLA
What it is
A JAX active-inference project for backprop-free world models and robotics benchmarks.
What I wanted to know
Could a robot choose actions by reducing expected surprise instead of only copying expert demonstrations?
What I tested
Generative state models, robotics benchmark adapters, and experiments around smolVLA/LeRobot-style embodied tasks.
What I learned
The framing made uncertainty easier to talk about, but only when the benchmark showed failures instead of hiding them in one score.
SystemRobotics certification
theplatform
Python, robotics benchmarks, safety scoring
What it is
A certification-oriented platform for evaluating robotic policies across safety, mechanical integrity, and mission success.
What I wanted to know
A robot score is more useful when you can see why it passed or failed.
What I tested
Scenario records, policy scoring dimensions, reporting structure, and certification-style pass/fail reasoning.
What I learned
One score is not enough. I need to be able to show a person what happened and why the robot passed or failed.
SystemTraining-data quality control
eidon-hand-detector
Python, CV, egocentric video QA
What it is
A computer-vision pipeline for checking whether robot-training POV videos contain useful hand presence, stability, lighting, and validity signals.
What I wanted to know
Bad training videos should get caught early instead of quietly hurting the model later.
What I tested
Hand detection, lighting checks, stability checks, validity flags, and review queues.
What I learned
Simple checks early in the pipeline make messy data much easier to clean.
PrototypeNeuroscience-inspired sequence dynamics
M.A.N.D.A
Python, Mojo, Apple Silicon, state-space models
What it is
A local playground for Mamba-style state-space models and faster experiments on Apple Silicon.
What I wanted to know
State-space models might be a better fit than attention-heavy models for some long-context local experiments.
What I tested
Python/Mojo implementation paths, Apple Silicon acceleration, and active-dynamics experiment scaffolding.
What I learned
Fast iteration changes the questions I am willing to try. A beautiful theory is not much fun if every experiment is painful to run.
PrototypeRobust digital watermarking
Python, Mojo, steganography, media processing
What it is
A hybrid steganography engine for embedding and recovering watermark-like information across image, audio, and text media.
What I wanted to know
A watermark is more useful if it can survive more than one media format.
What I tested
Payload encoding, media-specific embedding, recovery, and speed-sensitive pieces.
What I learned
The hard part is balancing strength and invisibility. One setting does not fit every file.
PrototypeEmergent physical rules
C++, cellular automata, simulation
What it is
A cellular-automata world where elements spread heat, interact, and produce physics-like behavior from local rules.
What I wanted to know
How much of a convincing little world can come from a small set of local rules?
What I tested
Grid updates, elemental interaction rules, temperature propagation, and behavior combinations.
What I learned
Rule order changes the whole feel of the world. I needed to watch it and measure it because either view alone could lie.
PrototypeModel mechanics from first principles
Rust, transformer internals
What it is
A compact Rust implementation of a decoder-only transformer designed to make attention and token prediction mechanics inspectable.
What I wanted to know
Building it in Rust forced me to actually understand the pieces instead of hiding behind a library.
What I tested
Token flow, attention blocks, residual connections, logits, and minimal inference structure.
What I learned
The idea is simple enough; the annoying details are tensor shapes and numerical correctness.
BenchmarkLocal speech recognition performance
Python, Distil-Whisper, Apple MPS
What it is
A benchmark comparing Distil-Whisper speech-recognition implementations on Apple Silicon MPS and faster-whisper paths.
What I wanted to know
Fast local transcription changes what kinds of voice tools are worth building.
What I tested
Different model versions and runtimes on a MacBook, with a focus on speed and how annoying each setup was to use.
What I learned
The runtime mattered as much as the model. Once transcription became fast enough, voice interfaces stopped feeling like a waiting room.
What it is
A personal AI doppelganger experiment built around conversational history and reflective simulation.
What I wanted to know
Would an assistant feel more like me if it could retrieve my actual conversational history, not just imitate my tone?
What I tested
Telegram data preparation, memory retrieval, persona prompting, and interactive conversation loops.
What I learned
The model is not the whole issue. Memory quality, consent, and weird identity questions show up fast.
SystemPersonal knowledge retrieval
Python, Streamlit, embeddings, Obsidian
What it is
A way to talk to my Obsidian vault while keeping my own notes as the source of truth.
What I wanted to know
A personal note vault becomes more useful when it can be queried conversationally instead of searched only by filename or keyword.
What I tested
Markdown ingestion, embedding retrieval, Streamlit UI, and answer generation over personal notes.
What I learned
A chat box cannot rescue badly organized notes. Retrieval made every weak title, giant note, and missing connection obvious.
PrototypeArchitecture reproduction
Python, PyTorch, transformers
What it is
A PyTorch implementation of a GPT-3-style transformer for learning how the model works.
What I wanted to know
Writing the model directly taught me more than only reading the paper.
What I tested
Transformer block composition, model configuration, training scaffolds, and paper-to-code translation.
What I learned
The architecture diagram is the easy part. Data, optimization, and compute are where the real system starts fighting back.
PrototypeGenerative modeling
Python, PyTorch, diffusion models
What it is
My implementation of a modified diffusion process so I could poke at the denoising loop directly.
What I wanted to know
What actually changes when I alter the noise schedule and some of the assumptions inside the diffusion process?
What I tested
Diffusion process structure, training loops, noise schedules, and generated sample behavior.
What I learned
Looking at samples was useful and also dangerously easy to fool myself with. Small changes needed a more consistent evaluation.
PaperSynthetic bio-molecular interfaces
HTML, scientific writing, biointerfaces
What it is
A speculative paper about tiny protein “antennas” that could make ultrasound-based neuromodulation more targeted.
What I wanted to know
Could tiny engineered bio-interfaces make ultrasound control more targeted and less invasive?
What I tested
The idea, paper structure, diagrams, and how clearly the argument reads.
What I learned
With speculative biology, clarity is the work. I had to separate what I was proposing from what is known and what is still completely unproven.