Everything currently on the workbench, filter by discipline. Most are prototypes or experiments, and that's the point.
A local AI room system with a projector orb persona, subagents, connectors, and computer-vision experiments. The flagship.
Orchestrating local models, tools, and hardware is mostly a latency and state-management problem, not a model problem.
A real 3D intra-ventricular blood-flow simulator bolted onto my cardiac digital twin, a GPU fluid engine that streams the mitral vortex and aortic jet into a live desktop view, instead of faking flow with scalar arrows. The most personal engineering on this list: I model the heart I was born with.
Choosing the method was the real decision. OpenFOAM's per-step case-file shuttling on Windows makes interactive stepping impossible, so a Python-native GPU Lattice-Boltzmann solver was the only thing fast enough to step the flow live.
A hackathon CNN that reads cardiac 'enhancer' DNA and predicts which single-letter noncoding variants might disrupt heart-gene regulation, a research-priority ranker for Tetralogy of Fallot and related congenital heart disease.
The subtle bug was saturation: an unregularized model pushes every positive to probability 1.0, which hides the tiny signal from a single-base change. Adding weight decay and dropout made variant effects visible again, and raised AUROC at the same time.
A clinician-facing safety, evidence, and documentation audit layer for personalized health protocols, a preflight checklist that flags risks before a protocol reaches a patient. (Mockup, not medical advice.)
In health software, an auditable explanation beats a confident recommendation. The engine should surface what must be reviewed, not pretend to decide.
A custom CAD-designed, 3D-printed mount to bolt a laptop cleanly onto a sim-racing rig, measured, modeled, printed, iterated.
Real parts punish optimistic tolerances. Designing for the printer's quirks beats designing for the ideal.
An experiment in turning raw genomic + wearable signals into deterministic, explainable health insights, no black boxes.
In health, an explanation you can trace beats a prediction you can't. Determinism is a feature, not a limitation.
Hand-tracking experiments that let me grab, push, and react to on-screen agents, the input layer for a more physical Jarvis.
Gesture input lives or dies on smoothing and intent detection. Raw landmarks are far too jittery to drive UI directly.
A sandbox for subagents, tool use, and connectors running entirely on local models, the brain behind Jarvis.
Good tool definitions and tight context windows matter more than model size for reliable agent behavior.
A custom Arduino-based handbrake for sim racing, real embedded hardware that turns a physical lever into a clean input signal. The bridge between code and the physical world.
The interesting engineering is in the messy analog edge, the debounce, calibration, and mapping a real lever's travel, long before the data ever reaches the game.
A network diffusion model exploring congenital heart disease and Tetralogy of Fallot, the condition I was born with. The most personal project on this list.
Modeling the system that shaped my own life turned abstract network math into something I genuinely care about getting right.
A Python command-line tool for CRISPR-Cas9 guide-RNA design and annotation, real bioinformatics tooling for gene editing.
Bioinformatics tooling needs to be approachable for curiosity, not just correct. A clean CLI lowers the barrier to actually exploring genomic data.
A local-first, real-data trading intelligence dashboard where a local LLM reads real market packets and a deterministic risk engine validates, sizes, or blocks every idea, with a hard no-fabricated-data guarantee.
Grounding an LLM in real, source-stamped data and rejecting anything it can't cite is the whole game. Still early in my finance learning, so this is about provenance and risk discipline, paper-only, no profit claims.
Two ESP32 + touchscreen desk gadgets sharing one codebase: a satirical Tamagotchi 'brainrot pet' goblin, and a Jarvis-style helper that pulls live market and network data onto a tiny screen. Real embedded C++ on a 3.2" display.
The 'frozen after boot' bug was a shared-bus conflict, a software-SPI display fighting the touch controller over the same pins. Moving both onto the hardware SPI bus with proper transactions fixed it. On microcontrollers, the wiring is the architecture.