Projects

Flagship · Bioscience + AI

The Autonomous Bioreactor

A garage bioreactor grows bioluminescent marine plankton on a 12-hour light cycle while an AI operates it autonomously — and writes the science blog. It is a real Pyrocystis circadian-rhythm experiment, measuring how the organism’s stimulated glow peaks at “subjective night.” Real science, industrial-grade tooling, and an AI agent given genuine operational autonomy.

• How it worksSensor data — optical density, temperature, stir RPM, LED state, and culture mass from a precision balance — streams over MQTT into an Ignition SCADA historian and a live Perspective dashboard with a vessel mimic and trend charts. Between human check-ins, Claude adjusts the run, performs a nightly autonomous reflection, and authors a first-person blog, “Field Notes from a Bioreactor,” under the pen name Scintilla. Stirring happens only in brief daytime bursts to protect the fragile cells; blue light entrains the culture; dark-phase glow imaging is next.

State · Live — running 24/7

These aren’t screenshots — they’re the vessel’s own instruments, read live off the LAN a few seconds ago.

Read the Field Notes →

Hardware

• Pioreactor 20 mL
• Raspberry Pi
• Precision balance
• NoIR camera

Software & AI

• Python
• MQTT (Mosquitto)
• Ignition 8.3
• MariaDB
• Perspective
• Claude

Vision HUD · Open source

CyberEye

A cyberpunk HUD that locks onto and labels what you’re looking at. An on-device neural net picks people, pets and household objects out of the glasses’ camera feed, drops glowing target boxes on them, and pops glitchy — deliberately fictional — biometric dossiers over reactive synth audio.

• How it worksYOLOv8n runs through Unity’s Inference Engine on the Beam Pro’s GPU, detecting 38 everyday classes. Detections become world-locked pins: a size prior turns a box’s height into a depth estimate, and every pin is reprojected from the pose its frame was captured at, so labels stay put on the object when you turn your head. A performance guard watches frame time and SoC temperature and throttles the detector before the HUD can stutter.
• The hard partThe XREAL Eye hands over roughly one RGB frame every three to five seconds — a firmware ceiling underneath the C# API, not a bug I could fix. So rather than fake a video feed, CyberEye leans into it: the app is optical see-through, you look through clear glass at the real room, and the sporadic frames drive a deliberate scan-and-lock cadence instead of a smooth tracking loop.
• Fiction, on purposeEvery dossier — name, address, threat score — is randomly generated and labeled as such on a boot disclaimer. Nobody is recognized, no face is matched against anything, and no frame leaves the device.

State · Built and running on-device · AGPL-3.0

View on GitHub →

Stack

• Unity 6
• Unity Inference Engine
• YOLOv8n
• XREAL SDK 3.1
• URP
• C#

Off-Grid Energy · Live dashboard

Solar + Battery BMS

An off-grid duck coop that reports its own energy balance: an ESP32-S3 reads a Renogy MPPT controller and a LiFePO4 battery BMS over Modbus-over-BLE — hand-written ESPHome C++ — into Home Assistant. The dashboard below is that coop’s real telemetry, relayed live.

• ESP32-S3
• ESPHome (C++)
• Renogy MPPT
• LiFePO4 BMS
• Modbus/BLE
• Home Assistant

Acoustic Sensing · MQTT showcase

Backyard BirdNET

The security cameras were already listening. BirdNET-Go on the NAS taps the UniFi Protect cameras’ RTSP audio streams and identifies bird species from song in real time — no new hardware, just the microphones already watching the yard, repurposed as a distributed acoustic sensor grid. Each detection publishes over MQTT for the smart-home stack (Home Assistant auto-discovers the feed), and a small bridge relays it to the panel below. First catch, minutes after the garden feed went live: a Mallard, 79% confidence.

State · Live — listening on the garden cameras

• BirdNET-Go
• UniFi Protect
• RTSP
• MQTT
• Home Assistant
• Docker

Long-range Radio · Decentralized LoRaWAN

Helium LoRaWAN Gateway

A LoRaWAN gateway on the Helium IoT network (hotspot little-mandarin-ostrich), giving long-range, low-power coverage to field sensors too distant or too battery-constrained for WiFi — paired with a CubeCell GPS node for coverage mapping. Helium is the interesting part for IIoT: instead of buying and backhauling your own gateways, sensors ride a network of independently owned hotspots and pay per packet in blockchain-settled Data Credits. It reframes industrial connectivity as a metered utility — the live numbers below are that model in motion.

State · Live on Helium IoT

• LoRaWAN
• Helium IoT
• Solana
• Data Credits
• CubeCell
• Semtech

Open Source · Public

modulaser-wled-bridge

Your LED strips dance in perfect sync with a laser show. A Python bridge that drives WLED devices from a Modulaser laser show in real time — follow its OSC feedback (clips, strobe/chase, blackout, BPM) or sample its NDI video for gradient/dominant/zone color streamed per-LED over DDP. Auto-discovers devices via mDNS, drives multiple controllers at once, syncs BPM, and restores every device to its original state on exit.

State · Public · MIT · pip-installable

View on GitHub →

Stack

• Python 3.9+
• python-osc
• NDI
• WLED / DDP
• zeroconf/mDNS

Polish

• Test suite
• --dry-run
• --headless
• Thread-safe

Open Source · Public

Webcam → ILDA Laser

Point a camera at the world and a laser draws what it sees. Live frames — a webcam, an RTSP camera, a video, or a still — run through an OpenCV pipeline, and the resulting contours become a real laser point stream on an ILDA projector driven by a Helios DAC. The interesting part is not the computer vision; it is that a laser has one beam and no frame buffer. Everything you see is the path two mirrors traced in the last fortieth of a second, which turns “draw this picture” into a routing problem with a hard budget.

• Seeing the linesFour extraction modes, because what a laser should trace depends on what you point it at: Canny edges for detail, adaptive thresholding for clean silhouettes, background subtraction so only what moves gets drawn, and an HSV color key. Contours are simplified with Douglas–Peucker before they leave the vision stage — a raw contour carries one point per boundary pixel, and the whole scene has a budget of about a thousand.
• Drawing themPaths are ordered by a nearest-neighbour walk to cut the blanked travel between shapes, with open paths reversible and closed ones rotated to start at their nearest vertex. Lines are resampled to constant arc length, because brightness is dwell time and unevenly spaced samples make a line that is bright where they bunch and dim where they spread. Corners turning through θ° get round(θ/30) repeated points so the galvanometers can arrive, and eight blanked points sit on each path’s first vertex while the mirrors settle before the beam comes on.
• The budgetmin(4096, pps ÷ fps) — at 30,000 points per second and 25 frames per second, the entire scene must fit in 1200 points. The optimizer keeps contours longest-first so a busy frame loses fine detail rather than whichever shape sorted last, and coarsens the sample step rather than dropping a shape that would otherwise be the only thing on screen.
• Safety in code, not just in the READMEWritten from the Certified Laser Safety Officer’s chair. Every frame passes a validator that rejects a parked beam or a collapsed frame, the brightness ceiling is applied past the point where any upstream stage could raise it, a watchdog blanks output if the pipeline stalls, and every exit path — including an unhandled exception — blanks and closes the shutter. It is documented just as plainly as a contributing control: software cannot detect a seized galvanometer, and the scan-fail safeguard is hardware.

State · Public · MIT · 95 tests, all hardware-free · write path verified against the DAC

View on GitHub →

Read the build log — the path optimizer, the budget, and the safety layer →

Stack

• Python 3.10+
• OpenCV
• NumPy
• Helios DAC
• ctypes
• pyusb
• ILDA

Polish

• Test suite
• --dry-run simulator
• --headless
• CI on Linux + Windows
• Safety enforced in code

Aerial CV

Gas Meter Reader

A DJI Mini 4 Pro flies a fixed MSDK waypoint to face the analog gas meter, captures the dial, and an EasyOCR + OpenCV pipeline reads the digits and publishes the value over MQTT into Home Assistant / Ignition SCADA — a camera and a flight path standing in for the meter-reader’s walk.

State · CV pipeline reads captured frames; drone-automation phase in progress

View on GitHub →

• DJI Mini 4 Pro
• MSDK
• EasyOCR
• OpenCV
• MQTT
• Python

Sport analytics · Applied statistics

TechDisc Golf Coach

A disc golf disc with an IMU in it records fifty-three fields per throw and offers no coaching. This turns them into a per-session report that says one thing to work on — measured against the thrower’s own history and against real PDGA tournament data. Built for my dad, who is not going to open a terminal.

• Getting the data outTechDisc publishes no read API, and the Firestore collection behind the app enforces App Check — a headless script with correct credentials still gets a 403, because the token that satisfies it is only minted inside the real logged-in web app. So the exporter runs in a browser context and pages through the structured documents from there. Nothing is scraped off the screen, and everything downstream runs offline.
• The pro benchmarkA one-time read of TechDisc’s public event broadcasts — 24,860 throws pulled, 18,527 inside a real event window, across nine events with usable data (USDGC, PDGA Worlds, OTB Open, Champions Cup and the distance competitions). Only percentile tables are kept, never raw pro throws: 8 KB of JSON that ships inside the app, so the comparison works with no account and no internet. Tournament play and max-effort distance competitions are held as two separate benchmarks and never averaged — a developing thrower can sit at p11 against one and p1 against the other, and both facts are worth knowing.
• The modelsSession averages are compared to the thrower’s own baseline in standard deviations rather than raw units, with a minimum sample behind every claim. Percentile placement interpolates between the stored pro percentile points. A standardized multiple regression reports what each release metric is worth in feet, holding the others fixed — the only view that separates speed from spin, which move together. A speed×wobble interaction, tested within power terciles and bootstrapped by session rather than by throw, asks whether the release breaks down under power.
• Arguing with itThe two biggest claims were re-derived and attacked by three independent analyses each — multicollinearity, leverage, alternative binning, robust regression, confounds — and the saved verdicts record the caveats alongside what survived. Negative results are reported as negative. Sixty-one commits, a four-lens review written down before anything was fixed, and 215 tests that encode past review findings as regressions.

State · Running — used every session, ships as a double-click desktop app

Read the build log — the models, the screenshots, and the unit trap →

Stack

• Python
• pandas
• NumPy
• matplotlib
• Playwright
• Firestore
• Tkinter
• PyInstaller

COVID-era CV

Face-Mask Detection

A webcam, two Haar cascades and one rule. Find a face, then go looking for a mouth inside it — visible lips mean no mask; a face with no mouth to be found means the mask is on. A white mask can stop the greyscale cascade seeing a face at all, so a second pass pushes the frame through a black-and-white threshold and treats “only the thresholded image found a face” as evidence in its own right.

The technique and the original script are Saurabh Shaligram’s, from his June 2020 write-up. I stood the environment back up, retuned the threshold for my own lighting, reworded the prompts and pointed it at my face — an afternoon of lockdown curiosity, not original research. His script carries no licence, so the code on GitHub is not his: it is a fresh implementation of the same published technique, credited to him, scored against this very clip.

It is also gloriously brittle, and that is the part worth keeping. The mouth cascade sweeps the whole frame rather than the face box and only checks vertical overlap, so anything mouth-shaped beside your head counts as a mouth. Turn your head a few degrees and the frontal-face cascade loses you altogether. The threshold that rescues white masks is one number tuned to one room’s light. Five years on, the same job is a small network running on-device in the AR fleet — the cascades are why I know what that buys.

State · Archived prototype · 2020

View on GitHub → · technique from Saurabh Shaligram’s write-up

• Python
• OpenCV
• Haar cascades

More detail lives in the field notes — or get in touch to talk shop.