Projects

Heavy Metals in Japanese Green Tea

Literature review · Risk modelling · Source‑backed numbers

I drink a lot of matcha and wanted to know if I should worry. Pulled twenty‑odd primary sources, normalized the units, and worked out which elements actually matter at realistic daily doses. The answer turns out to be more nuanced than either the alarmist or the dismissive crowd suggests.

• Aluminium, manganese, lead, cadmium, fluoride: ranges and dose context
• Brewing physics: which metals leach into the cup vs. stay in the leaf
• Comparison against EFSA, WHO, and US reference values

Write‑up forthcoming.

Extracting Sprites From a 1993 Game

Reverse engineering · C++ · Pillow · 480 sprite tiles

A pipeline that pulls every graphical asset out of Per.Oxyd's proprietary .DAT archive, slices the sprite sheets into a clean 32×32 grid, and emits separate 1‑bit transparency masks for use in a modern engine. A direct sequel to the de‑dithering CNN, and feedstock for the maze prototype below.

• Wraps the open‑source Enigma C++ oxydlib parser
• Auto‑sorts backgrounds, UI, and 480 individual tiles into folders
• One‑shot Python orchestrator: download, compile, extract, clean up

Write‑up forthcoming.

A Hide‑and‑Seek Labyrinth Game

LÖVE · Lua · Box2D · Procedural mazes

A game I sketched in 2017 and finally started building: navigate a maze of monsters with mines, cameras, and keyhole peeks. Per‑level mazes generated with chamber‑aware DFS and rounded subtractive corners; physics, audio, and key/lock layers tuned by config. Uses the sprites extracted above.

• Box2D physics with rounded corner collision geometry
• Graph‑distance debug overlays for shortcut analysis (F3)
• Detached audio that survives level reloads

Prototype, no public build yet.

Finding Hotels Without Google Maps

Open data · GTFS routing · Multi‑API pipeline

A hotel‑search pipeline built entirely on open data: OpenStreetMap for locations, MTA’s GTFS feed for schedule‑aware subway routing, OSRM for walking distances. Found 316 hotels reachable within 20 minutes, no Google APIs required.

• Dijkstra‑like GTFS propagation with real departure times
• OSRM batch routing for walking distance calculations
• Interactive scatterplot of price vs. value

Training an LLM Brand Advisor

LoRA fine‑tuning · ONNX quantization · In‑browser inference

Fine‑tuned Gemma 3 1B into a brand‑loyal car advisor using LoRA, then shipped the quantized model to the browser via ONNX and WebGPU. Training to deployment in one pipeline.

• LoRA rank 8 on attention matrices, trained on 50 examples
• Response‑only loss masking for focused gradient updates
• 5 GB → 1.3 GB via INT8 quantization, runs in‑browser

A Search Engine for My Own Life

Semantic search · HNSW · Fully offline

Semantic search over years of AI conversations, audio logs, diary entries, and Discord messages. Sentence‑transformers and HNSW for instant queries. No data leaves the machine.

• HNSW approximate nearest‑neighbour search via hnswlib
• Configurable chunking per document type
• Incremental indexing: only new content gets processed

An E‑Paper Dashboard for the Hallway

Raspberry Pi · Ollama · 12.48″ e‑ink

An ambient display showing AI‑curated news, gold prices, and my calendar, powered by a Raspberry Pi running a unified scheduler that orchestrates RSS collection, embedding‑based filtering, and LLM curation. No backlight, no notifications.

• News pipeline: RSS → BGE‑M3 embeddings → Ollama curation → HTML
• Central scheduler with per‑task intervals and state persistence
• Deployed via SSH/SFTP with systemd service management

A Podcast Transcription Pipeline

ASR · Speaker diarization · LLM post‑processing

A local pipeline that chains NeMo Parakeet for speech recognition, pyannote for speaker identification, and Ollama for cleanup. Produces speaker‑labelled transcripts from any audio file.

• Chunked transcription with Levenshtein‑based overlap merging
• Word‑level timestamps aligned to diarization segments
• LLM post‑processing for speaker naming and correction

Building a Personal AI Newspaper

Embeddings · Vector calibration · LLM curation

An automated news‑curation pipeline that filters RSS feeds using embedding‑based similarity matching against personal interests, with careful calibration to handle AND‑combined topic constraints.

• Solved single‑concept vs multi‑concept bias via centroid scoring
• Contrastive lift gating to reject random‑looking matches
• Final curation by Gemini, deployed to a Raspberry Pi

Rating Photos by Aesthetic Quality

AesCLIP · Offline scoring · CLI app

Scores photos 0–10 for aesthetic quality using AesCLIP, entirely offline. Processes a full photo archive incrementally, persists scores in SQLite, and includes a slideshow for browsing the best results.

• AesCLIP (ACM MM 2023) with CLIP ViT‑B/16
• Incremental processing: skips unchanged files via size + mtime
• Production CLI with Typer, Rich output, fail‑fast errors

Letting an LLM Name My Files

Document automation · Ollama · Metadata extraction

Batch‑renames academic papers and invoices by extracting metadata with a local LLM. Papers become “[year] [author] - [title]”, invoices become “[date] [institution] - [title]”.

• Supports PDF, EPUB, XPS, CBZ, and FB2 via PyMuPDF
• TOML‑configurable prompts for different document types
• Duplicate‑aware naming with conflict resolution

Painting With Noise Fields

Stable Diffusion · CLIP · Spatial control · Gradio

Explores what happens when noise in Stable Diffusion carries spatial meaning, with fixed concepts in some regions and free morphing in others. CLIP embeddings with spatial masks for per‑location control over generation.

• Three noise dimensions: visual, concept (512‑dim), feature (768‑dim)
• Spatial masks control where concepts are fixed vs. variable
• Gradio interface for interactive mask painting

Growing Mazes From Images

OpenCV · Generative art · SVG

Maze walls that follow the contours of an image. Gradient fields guide the maze topology so passages flow along edges and textures, producing mazes that look like they grew from the picture.

• Scharr gradients → angle fields → Floyd‑Steinberg dithering → SVG
• Square and triangle grid support with per‑cell orientation
• Tunable loopiness, branchiness, and dead‑end rate

Mazes as Spectral Graphs

Spectral graph theory · Eigenvalues · Tensor factorization

Applying spectral methods to maze design. The Laplacian eigenvectors reveal a maze’s fundamental topology, and changing eigenvalues changes how a maze feels without touching individual walls.

• Normalized Laplacian and spectral embedding for maze topology
• Interactive eigenvalue control via ipywidgets sliders
• Non‑negative PARAFAC tensor factorization for decomposition

Removing Music From Audiobooks

YAMNet · Audio classification · Segment removal

Uses Google’s YAMNet to classify audio segments as speech or music, then surgically removes background music from audiobooks. Also merges multi‑chapter M4A files with preserved metadata.

• Frame‑by‑frame speech vs. music scoring via TensorFlow Hub
• Minimum segment length filtering to avoid false positives
• FFmpeg normalization pipeline for consistent output

Finding Similar Books With Embeddings

Universal Sentence Encoder · Book similarity · NLP

Embeds entire ebooks into 512‑dimensional vector space using Universal Sentence Encoder. Similarity search, style matching, categorization, and duplicate detection across large libraries.

• Samples 120 sentences from mid‑book to capture voice, not front matter
• Median embedding per book for noise‑robust representation
• Gaussian models for genre categorization

Sequences That Look Random to Humans

Cognitive science · Algorithm design · Perception

True randomness feels wrong to humans; we see patterns in noise. This algorithm generates sequences that actively avoid the structures our brains detect: mirrors, runs, and repetitions.

• Exponential penalties with distance‑based mitigation
• Tuned per pattern type: repetition, mirror, monotonic run
• Output rated as “more random” than actual random sequences

Teaching a Network to See Past Dithering

CNN · Retro game restoration · Custom training pipeline

Built a CNN to convert 4‑bit dithered textures from the retro game Per.Oxyd back to full 24‑bit colour. Custom training data from Floyd‑Steinberg dithering, with augmentation to handle checkerboard shadow artifacts.

• ESPCN‑based architecture, 5 conv layers, trained on 2,500 photos
• Checkerboard shadow augmentation to prevent false de‑dithering
• 30 dB PSNR, 63 hours of documented iteration

Real‑Time Video Upscaling With ESPCN

Super‑resolution · PyTorch · Sub‑pixel convolution

A PyTorch implementation of the ESPCN super‑resolution network, upscaling low‑resolution video without the smeared look of bicubic interpolation. Image and video input with standard benchmark evaluation.

• Sub‑pixel convolution moves upscaling to the end for efficiency
• Solved YCbCr colour shift by switching to RGB training
• ~31.5 dB PSNR on Set5, Set14, BSD100

Simulating Grapheme‑Colour Synesthesia

Color science · Cognitive perception · PDF generation

A simulation of what text looks like to a grapheme‑colour synesthete. Each letter has a fixed colour, but neighbouring letters blend and influence each other through Gaussian‑weighted fields, modelling the perceptual experience, not just the mapping.

• Cassidy Curtis colour mapping with Gaussian neighbour influence
• HSL polar mixing for perceptually uniform blending
• Outputs coloured PDF via ReportLab

Reverse Engineering an E‑Bike Controller

PIC Assembly · Firmware analysis · Embedded systems

Disassembled and analyzed PIC16F946 firmware from a TranzX e‑bike motor controller. Traced motor PWM, battery monitoring, pedal‑assist logic, and display communication across multiple firmware versions.

• Multiple firmware hex dumps analyzed (SS606, SU806, AT8U3)
• Created instruction set and register reference tables
• Hardware analysis with EAGLE schematics