I build hardware that actually ships.
And software that blends in.

• 42 landmarks tracked across both hands in real time, from a standard webcam.
• 60 fps end to end with sub-50 ms hand-to-render latency on consumer hardware.
• Pinch to grab, scale and rotate. Air-draw in space and extrude sketches into 3D models.
• STL, GLB and GLTF support: inspect real assets with no headset, no controllers, no install.

Inspired by Tony Stark’s workshop and the 2013 SpaceX gesture-CAD demo, rebuilt with modern AI so it works for anyone with a browser. It evolved from laggy shape manipulation into a full spatial interface — and then into Mexger, where the gestures finally drive real parametric geometry.

Two-brain architecture: a Raspberry Pi CM5 handles vision and on-device AI while an ESP32-S3 manages sensor I/O and power, driving a reflective-waveguide see-through display. The AR and firmware modules are running in bench tests.

In development for V2: a fully on-device assistant (quantised Phi-3-mini, whisper.cpp speech recognition, Piper TTS and retrieval memory) and continuous two-point ECG through ear-tip electrodes. No cloud in the loop.

The first pair: a fully soldered, daily-driver prototype with an ESP32 driving a 1.51-inch transparent OLED over SPI, with audio, capacitive-touch controls and health sensing no off-the-shelf prototype offered at the time — heart rate, SpO₂ and IR skin temperature, from a pair of glasses. Every lesson from V1 is a design decision in V2.

Hips and knees actuated by harmonic-drive BLDC motors on a rigid frame running from waist to foot, crutch-supported, with simple commands for stand, sit and step. Safety leads the control design: controlled descent on any failure, conservative lean limits, hard interlocks.

Validated so far in a 12,000-run Monte-Carlo sizing sweep and a browser physics sim passing 8 of 8 stability tests. Devices like this exist today at the price of a car; the goal is a design the people who need it can actually reach. Next milestone: a single-joint bench rig under real load.

A fully soldered ESP32 device on a 320×240 TFT reporting nine measurements: heart rate two independent ways (ECG R-peaks and PPG, cross-checked), HRV (RMSSD, SDNN, pNN50), respiration rate, SpO₂, perfusion index, FOV-corrected surface temperature, relative alcohol response and supply-rail voltage.

Acquisition and rendering are split across both ESP32 cores; the ECG task holds a fixed 250 Hz tick and never touches a blocking bus. R-peaks are detected with Pan-Tompkins plus parabolic interpolation for sub-sample timing.

Every on-screen value is measured or shown as “--”: SpO₂ is flagged as an estimate until calibrated, and alcohol is reported as a sensor resistance ratio, not BAC. The ECG channel was compared against a hospital report.

Built strictly from on-hand components on a TVS Racing shell, no scope creep allowed. An ESP32-WROOM-32 drives an SSD1306 through a 25 mm beam-splitter prism with a mirrored render, an MPU6050 supplies attitude, and phone navigation arrives over BLE, auto-swapping the HUD between a 24-hour clock and live Google Maps directions.

Servo-actuated visor, integrated audio, health-monitoring hooks and dual I²C buses round it out. Prism render, IMU attitude, BLE navigation and the servo visor are all working on hardware.

Discover and install models and community skills across Claude, ChatGPT, Gemini, Perplexity, Cursor and more. PromptArm rewrites raw prompts through 8 modes for 16 models from 7 providers, so you see how the same idea lands everywhere. A Library tracks token usage and quota-refresh timers per provider; creators publish their own skills, prompt packs and agents to the store.

Bring your own API keys, encrypted with AES-256-GCM, or start on the free Llama 3.3 70B tier. Prompt content is never stored; export or delete your data any time.

A session timer converts focus into G-Score, with streaks, achievements, a socially-earned Aura and grind levels from Beginner to Legend. Lock-in mode lets you stake your G on finishing a session; Coffee Pools bring the stakes to groups.

Audited and shipped in about seven weeks: 14,270 lines, 21 Supabase tables with row-level security on every one, realtime feeds, and a campus QR launch at Queen’s University Belfast.

A WhatsApp-style app where you chat with AI versions of specific people. Describe them, or import a real chat export and it picks up their voice on its own. Every message is stored and older threads fold into running summaries, so personas remember you across devices and days.

The clever bit: the voice keeps learning, but only from the persona’s side of the conversation, never yours. It stays them; it never drifts into mirroring you. Groq for near-instant replies, Supabase Postgres for durable memory, Flask serverless on Vercel.

Occupation is the most underused signal in dating: a proxy for lifestyle, schedule, ambition and shared grind. Every app lets people judge by job quietly; Belovrr makes it the explicit foundation, scored through an Occupation Compatibility Score.

No swiping. A vertical feed with hold-and-slide up to like, down to pass, around a radial hub, and a tap-to-reveal glassmorphism profile. Landing page and waitlist are live and capturing emails; a working app prototype is built.