Experiments, prototypes, notes

RESEARCH

Research-shaped work from repo evidence: latent world models, active inference, robotics evaluation, local speech systems, model internals, simulation, and provenance.

Research Map

9
Prototype
3
System
1
Benchmark
1
Paper
Portfolio Research

Themes

Latent world models for manipulation Active inference for robot control Robotics certification Training-data quality control Neuroscience-inspired sequence dynamics Robust digital watermarking Emergent physical rules Model mechanics from first principles Local speech recognition performance Personal memory systems Personal knowledge retrieval Architecture reproduction Generative modeling Synthetic bio-molecular interfaces
Prototype Latent world models for manipulation

Xlab-Jepa

Python, V-JEPA/V-JEPA2, SO-101 robotics

Abstract

A robotics research scaffold for testing whether latent video representations can support action-conditioned prediction and planning for manipulation tasks.

Hypothesis

A compact JEPA-style latent state can preserve enough task-relevant structure to guide planning without reconstructing every pixel.

Tested

Robot-video ingestion, representation checkpoints, action-conditioned dynamics loops, and SO-101 adaptation paths.

Discovered

The highest-leverage engineering work is not the model alone; data quality, action alignment, and evaluation harnesses decide whether latent planning is measurable.

Prototype Active inference for robot control

taif

JAX, active inference, LeRobot, smolVLA

Abstract

A JAX active-inference framework exploring backprop-free generative world models and robotics benchmarks.

Hypothesis

Policy selection can be framed as minimizing expected prediction error or free-energy-like objectives instead of only imitating expert actions.

Tested

Generative state models, robotics benchmark adapters, and experiments around smolVLA/LeRobot-style embodied tasks.

Discovered

Active-inference framing is promising for interpretability, but benchmark design must make uncertainty and failure modes visible.

System Robotics certification

theplatform

Python, robotics benchmarks, safety scoring

Abstract

A certification-oriented platform for evaluating robotic policies across safety, mechanical integrity, and mission success.

Hypothesis

Robotic policy evaluation becomes more useful when it produces audit-style evidence instead of isolated demo scores.

Tested

Scenario records, policy scoring dimensions, reporting structure, and certification-style pass/fail reasoning.

Discovered

Safety evaluation needs human-readable trace artifacts; aggregate metrics alone do not explain why a policy should be trusted.

System Training-data quality control

eidon-hand-detector

Python, CV, egocentric video QA

Abstract

A computer-vision pipeline for checking whether robot-training POV videos contain useful hand presence, stability, lighting, and validity signals.

Hypothesis

Filtering low-quality demonstration videos before training improves downstream robot learning more cheaply than trying to fix noisy data later.

Tested

Hand detection, lighting heuristics, stability checks, validity flags, and dataset triage workflows.

Discovered

Simple quality gates are powerful when they are placed early; they turn ambiguous dataset problems into actionable review queues.

Prototype Neuroscience-inspired sequence dynamics

M.A.N.D.A

Python, Mojo, Apple Silicon, state-space models

Abstract

A research prototype around Mamba-style active dynamics, state-space modeling, and accelerated local experimentation.

Hypothesis

State-space sequence models can offer a useful alternative to attention-heavy architectures for long-context dynamics and efficient local execution.

Tested

Python/Mojo implementation paths, Apple Silicon acceleration, and active-dynamics experiment scaffolding.

Discovered

Implementation ergonomics matter as much as theory when experimenting locally; fast iteration changes what can be investigated.

Prototype Robust digital watermarking

StegaDNA

Python, Mojo, steganography, media processing

Abstract

A hybrid steganography engine for embedding and recovering watermark-like information across image, audio, and text media.

Hypothesis

Cross-media encoding strategies can make provenance signals more robust than single-format watermarking.

Tested

Payload encoding, media-specific embedding paths, recovery workflows, and performance-sensitive components.

Discovered

Robustness and perceptual invisibility are in tension; useful systems need explicit tradeoff controls rather than one universal embedding recipe.

Prototype Emergent physical rules

ATLA

C++, cellular automata, simulation

Abstract

A cellular-automata simulation engine for elemental interactions, heat transfer, and rule-driven physical behavior.

Hypothesis

A small set of local rules can create expressive emergent behavior suitable for interactive simulation and game-like worlds.

Tested

Grid updates, elemental interaction rules, temperature propagation, and behavior combinations.

Discovered

Rule ordering and conservation assumptions shape the perceived physics; debugging emergence requires visual and numerical feedback together.

Prototype Model mechanics from first principles

handmade_transformer

Rust, transformer internals

Abstract

A compact Rust implementation of a decoder-only transformer designed to make attention and token prediction mechanics inspectable.

Hypothesis

Rebuilding a transformer in a low-level language clarifies architecture details that framework abstractions hide.

Tested

Token flow, attention blocks, residual connections, logits, and minimal inference structure.

Discovered

The core architecture is conceptually small, but correctness depends on careful tensor shape discipline and numerically boring details.

Benchmark Local speech recognition performance

Distil-Whisper-Test-mps

Python, Distil-Whisper, Apple MPS

Abstract

A benchmark comparing Distil-Whisper speech-recognition implementations on Apple Silicon MPS and faster-whisper paths.

Hypothesis

Apple Silicon acceleration can make local speech transcription practical enough for real-time assistant workflows.

Tested

Model variants, runtime backends, inference speed, and implementation ergonomics on a MacBook-class device.

Discovered

Backend choice can dominate user experience; faster local transcription unlocks interaction designs that cloud-only latency makes awkward.

Prototype Personal memory systems

openai-hackathon

Python, Telegram exports, LLM personas

Abstract

A personal AI doppelganger experiment built around conversational history and reflective simulation.

Hypothesis

A personal archive can support a more grounded assistant persona when retrieval and conversational style are both modeled.

Tested

Telegram data preparation, memory retrieval, persona prompting, and interactive conversation loops.

Discovered

Personal AI systems quickly become social and ethical systems; memory quality, consent, and misrepresentation matter as much as model quality.

System Personal knowledge retrieval

RAGonOBsidian

Python, Streamlit, embeddings, Obsidian

Abstract

A retrieval app for chatting with an Obsidian vault using local notes as context.

Hypothesis

A personal note vault becomes more useful when it can be queried conversationally instead of searched only by filename or keyword.

Tested

Markdown ingestion, embedding retrieval, Streamlit UI, and answer generation over personal notes.

Discovered

Good answers depend on chunking and note hygiene; retrieval makes hidden structure in a personal archive very obvious.

Prototype Architecture reproduction

GPT-3 paper implementation

Python, PyTorch, transformers

Abstract

A PyTorch implementation of GPT-3-style transformer architecture for language-model study.

Hypothesis

Implementing the architecture directly is the fastest way to understand scaling-era transformer design choices.

Tested

Transformer block composition, model configuration, training scaffolds, and paper-to-code translation.

Discovered

Paper reproduction exposes the gap between architectural diagrams and training reality: data, optimization, and compute dominate the practical system.

Prototype Generative modeling

Modified DDP

Python, PyTorch, diffusion models

Abstract

An implementation of a modified deep diffusion process model for generative modeling experiments.

Hypothesis

Changing diffusion process details can reveal how denoising schedules and model assumptions affect generation quality.

Tested

Diffusion process structure, training loops, noise schedules, and generated sample behavior.

Discovered

Small algorithmic changes are hard to judge without strong evaluation rituals; visual inspection is useful but insufficient.

Paper Synthetic bio-molecular interfaces

av-sonogenetics-paper

HTML, scientific writing, biointerfaces

Abstract

A paper page exploring logic-gated protein mini-antennas for ultrasound-addressable neuromodulation and readout.

Hypothesis

Engineered molecular interfaces could support selective, noninvasive ultrasound interaction with neural systems.

Tested

Conceptual architecture, paper structure, explanatory diagrams, and scientific argumentation flow.

Discovered

For speculative biointerface work, clarity of assumptions is the product; readers need to see what is proposed, what is known, and what remains unproven.