Things I build outside work. By day I am a semiconductor process and materials engineer; evenings and weekends produce what's in this repo: thermodynamic screening on open materials data, materials-characterization pipelines (SEM super-resolution, impedance/DRT), local-first AI tooling, parametric CAD written as Python, medical-imaging pipelines, and an espresso machine that files telemetry reports.
In a hurry? Three artifacts carry the flavor: the caught-and-fixed API data bug in the materials screening, the NIIMBOT driver's GOTCHAS.md (identical printers, different firmware dialects), and the GMA pipeline's honest scorer that declares three subscales NOT_COMPUTABLE rather than guessing.
Everything here shares three habits:
- Code over artifacts. Parts, pipelines, and reports are all regenerable from scripts. No orphaned binaries; no personal data ships with this repo, ever.
- Validation as a first-class step. Watertight checks and containment truth tables for printed parts, synthetic-data test suites for pipelines, adversarial review passes for research documents.
- Calibrated AI-assist notes. All of this was built with Claude in the loop. Each project documents concretely what AI acceleration bought and where it confidently failed, because both halves are the interesting part.
I'm a process and materials engineer by day, but everything in this repository is a personal, nights-and-weekends project. The materials-science, chemistry, and physics explorations here are pursued purely out of curiosity, on my own time. They use only publicly available data and open source codes, produce only self-generated results, and are completely unrelated to my current daily activities at my current employer - no proprietary data, no proprietary systems, no affiliation.
| Section | What's inside |
|---|---|
| ai-tooling/ | Two local-first tools over your own paper library: a Zotero 7 plugin doing on-device retrieval (BM25 + ONNX embeddings) with only the most relevant passages sent to Claude, and a fully offline Ollama batch summarizer that turned ~3,900 PDFs into a searchable Obsidian idea bank |
| advanced-materials-characterization/ | Two characterization examples: SEM micrograph super-resolution (classical vs deep-learning upscaling on Apple Silicon) and an EIS/DRT toolkit that deconvolves impedance spectra into relaxation-time peaks; both run on bundled synthetic data |
| audio/ | A construction-noise monitor born of a napping baby vs. the jackhammers next door: camera clips to a fully local dBFS timeline, noise-event detection, and an hour-bucketed report of loud minutes outside permitted hours (levels only, no speech, ever) |
| claude-skills/ | Nine working Claude Code skills: parametric CAD, academic figures, arXiv PDFs, hybrid retrieval, research orchestration, transcript tooling |
| computational-materials/ | Materials Project interface-reaction screening (gate-dielectric stability, an 18-gas etch-chemistry study at 0 K vs 300 K, a caught-and-fixed API data bug), plus a scriptable SRIM/TRIM wrapper that drives ion-transport sputtering simulations across a parameter grid, headless in Docker |
| espresso-gaggiuino/ | An open-source machine mod run as a data project: 639-shot telemetry corpus, drift detection, troubleshooting as differential diagnosis |
| hardware-tools/ | Bluetooth batch printing on macOS for a NIIMBOT label printer that ships phone-app-only: reverse-engineered BLE protocol, firmware-dialect detection, dry-run byte tracing |
| medical-imaging/ | An infant-movement video pipeline (SAM 3 on Apple Silicon) and a DICOM MRI toolkit, both de-identified and code-only, with caveats and shortcomings documented |
| 3d-printing/ | Seventeen parametric builds plus two design playbooks; every part is a Python or OpenSCAD program |
pySRIM sputtering & ion-transport - where ions stop in an optical coating: projected range and straggle building up frame by frame (schematic; the real numbers come from TRIM in Docker)
 |
 |
| SEM super-resolution classical vs deep-learning x4 synthetic sample, Apple Silicon |
EIS to DRT one Nyquist arc into peaks validated on known tau |
 |
 |
| Gate-dielectric screening the O-Si-Ta convex hull: Ta2O5 + Si -> silicides + SiO2 |
GMA video pipeline per-keypoint symmetry screen in-norm limbs blue, flagged red |
 |
 |
| DICOM MRI toolkit annotating anatomy in code shown: public MNI152 template |
Kumiko lattice vent cover parametric asanoha pattern bed-split with real joinery |
A Zotero 7 plugin that answers questions about your own paper library from inside Zotero: a section-aware BM25 index and a quantized ONNX embedding model (bge-small-en-v1.5, running in a ChromeWorker via Transformers.js) retrieve on-device, Reciprocal Rank Fusion merges the rankings, and only the most relevant passages go to the Claude API. Papers never leave the machine at index time; the network surface is one hostname. TypeScript, esbuild, CI that deliberately builds without the model weights to keep the degraded path tested. Has an ARCHITECTURE.md with dataflow diagrams.
The offline sibling in this section, an Ollama paper summarizer, takes the opposite tradeoff: it batch-summarizes a whole PDF library into structured Markdown with a local Ollama model - nothing sent anywhere, no API key - so the summaries become a greppable idea bank an Obsidian similarity plugin surfaces while you write. It has summarized about 3,900 papers entirely offline; a carefully constrained prompt keeps every note the same shape instead of a random abstract.
Two example projects that look harder at a material, both running on bundled synthetic data so nothing real ships. A SEM super-resolution tool batch-upscales grayscale electron-microscopy images on Apple Silicon, putting a five-model deep-learning zoo (Real-ESRGAN, SwinIR, HAT, Swin2SR, BSRGAN) next to a classical Lanczos baseline so you can decide per image whether the transformer actually resolves detail or just costs 100x more compute to look the same. Its companion, an EIS/DRT toolkit, deconvolves an electrochemical impedance spectrum into a Distribution of Relaxation Times: one blurry Nyquist arc becomes a comb of peaks along a timescale axis, so you can count how many processes are actually hiding in it. It ships headless (pyimpspec TR-RBF, CSV out) and as a Streamlit slider app, and validates itself on a synthetic Randles + two-ZARC spectrum with known time constants rather than asserting an answer. Both tools document the same class of AI-assist failure caught in review: a missing import that would crash on the first run, and prototypes that hardcode a local absolute data path.
Prolonged construction next door and a napping infant raised one question that opinions could not settle: how loud, how often, and inside or outside working hours? A noise monitor answers it with data. It extracts the audio track from home-camera clips (ffmpeg), builds a windowed dBFS level timeline, detects noise events against a rolling-median baseline with hysteresis, renders spectrograms, and produces an hour-bucketed report flagging loud minutes outside a permitted-hours window. It is fully local (numpy, no cloud, no API keys) and processes no speech at all - noise levels and timing only, deliberately nowhere near what anyone said. Levels are honest relative dBFS from an uncalibrated camera microphone, not dB(A). Ships with a seeded synthetic demo so it runs without any recording.
Nine of the skills that automate the rest of this repo, published as adaptable public copies: the parametric-CAD skill that built the 3d-printing section, hybrid Zotero retrieval, deep-research orchestration into Obsidian, YouTube transcript tooling, arXiv-style PDF generation, and a coupon tester whose guardrails (never purchase, never touch payment fields) are the interesting part. Skills with private local dependencies declare them honestly in marked adaptation notes.
A Materials Project interface-reaction tool that walks pseudo-binary mixing lines for reaction-energy kinks, MP-website style. The demo reproduces the classic gate-dielectric screening from open data: Ta2O5 decomposes against silicon into silicides + SiO2 (why it lost the high-k race) while HfO2 and Si3N4 sit at exactly 0.00 eV/atom (why they get to touch silicon). The companion study runs silicon against 18 fab gases - Bosch etch, chamber cleans, MEMS release - at 0 K and at 300 K via a SISSO-Gibbs mode, measuring exactly when the 0 K ranking stops being trustworthy and flagging where the 300 K mode itself deserves skepticism. Also documents a silently changed API default (mixed GGA/R2SCAN hull) that was corrupting results until a literature cross-check caught it.
The second tool, pySRIM, wraps SRIM/TRIM - the standard ion-transport Monte-Carlo, normally a click-heavy Windows GUI - so it runs headless in Docker (Wine + xvfb, --platform linux/amd64 on Apple Silicon) across a whole parameter grid. It sputters an optical thin-film stack - low-index SiO2, high-index TiO2, and the ITO transparent conductor - and returns projected range, straggle, vacancy/damage profiles, sputter yield, and energy partitioning. The honest-notes section is the point: it says where SRIM is trustworthy (range, damage) and where it is not (absolute low-energy sputter yields), and documents the import-time monkey-patch that keeps a decade-old package alive on modern PyYAML.
Replaced the controls of a Gaggia Classic Evo Pro with the open-source Gaggiuino controller (STM32 + ESP32, pressure/flow profiling, PID), then treated eight months of daily use as a data project: a 639-shot telemetry corpus pulled from the machine's HTTP API, a machine-drift analysis that isolated OPV spring fatigue from grinder and coffee variables, and troubleshooting docs structured as differential diagnoses with explicit priors. Safety claims were checked against the firmware source, not forum consensus.
Batch label printing for the NIIMBOT D110, which ships with no computer software at all - phone app only, one label at a time. I liked the printer itself, but tapping labels out one at a time on a phone was never going to knock out the good-husband weekend chores: a week of baby-food containers, a whole-pantry reorganization. This CLI takes a text file and prints the entire stack over one BLE connection. Under it sits a reverse-engineered protocol driver whose best story is in its GOTCHAS file: visually identical printers ship with different firmware dialects, and the _M variant returns byte-perfect success traces while printing blanks. Folder-licensed GPL-3.0 because one upstream protocol source is GPL - the one non-MIT corner of the repo, labeled as such.
Two side quests built to understand clinical imaging from first principles instead of passively receiving it. Neither ships any real imaging or reports (each carries one explicitly labeled synthetic or public-template example), and neither is remotely a medical device.
- GMA video pipeline: turns consumer phone video of an infant into segmentation masks (SAM 3 run deliberately on the Apple Silicon GPU through PyTorch's Metal/MPS backend, via a five-patch compatibility shim), pose keypoints, kinematic features from the published GMA computer-vision literature, and a deliberately honest proxy score that reports what it cannot see as NOT_COMPUTABLE rather than guessing.
- DICOM MRI toolkit: from a hospital CD to an interactive 3D reconstruction: series matching by pulse parameters, longitudinal difference maps, spline-curvature candidate detection on ventricle walls, five-sequence cross-validation, and a self-contained clickable HTML report.
Each has an ARCHITECTURE.md with system diagrams.
Every part is a Python or OpenSCAD program; every STL is reproducible by running a script. Seventeen curated builds from a workshop of about 30, printed on a Bambu Lab H2D: a six-version rice-bowl gravity dispenser, a child-safe nursery floor register whose safety constraint is an assert (and its wet-room and maximum-airflow descendants), a print-in-place sliding-damper register whose slider is collision-swept across its full travel, a 90-degree louver diverter designed by free-area arithmetic, a kumiko-lattice vent cover split for the bed with real joinery, and a multi-start replacement thread nut bracketed by an overnight eight-ring test matrix. Start with the parametric design gotchas playbook: cross-project design rules, each traceable to a printed mistake.
Nearly everything here was pair-built with Claude Code. The per-project "AI-assisted build notes" sections are deliberately specific: the MPS shim and validation harnesses that AI made fast, and also the silently wrong jerk normalization, the alignment approach that was confidently wrong twice, the fillet that no-opped five versions, and the ten citation errors an adversarial review pass caught. The stable division of labor across every section: AI writes geometry, plumbing, and checks; the human owns physical judgment, spatial semantics, and looking at the picture.
MIT for all code (see LICENSE), with one labeled exception: hardware-tools/niimbot-labelmaker/ is GPL-3.0 because part of its protocol knowledge derives from a GPL-3.0 upstream. Documentation and research notes may be quoted with attribution. One 3D printing project adapts a community concept and credits it in its README.