NeuroAccelerator — Personalised Health Intelligence
Most health apps collect data. NeuroAccelerator analyses it. Using advanced signal decomposition methods drawn from medical device algorithm development, the app mathematically reveals hidden relationships between a user's daily activities and their symptoms — surfacing patterns that are invisible to intuition alone.
The analytics plot above shows the cyclic signature of weekend partying on sleep quality, fatigue, and diet — automatically extracted from logged behaviour across weeks of data. Each axis is an independent component: a latent behavioural dimension discovered from the user's own data, not a predetermined category. The colour gradient from blue to red traces time, making the repeating weekly cycle immediately visible.
Users control their own depth of analysis. At minimum, logging diet and symptoms reveals food-to-symptom relationships mathematically. Those who wish to go further can unlock analytics for energy and mental health patterns, relationship health, and more — each layer adding resolution to the picture of what drives how they feel.
Typing Mastery Profiler
Most typing tutors measure speed. This tool measures the causes of speed. Each practice session is plotted as a point in a 3D profile space: keystroke regularity (consistency of timing between key presses), characters per minute, and hand independence (how freely each hand moves without slowing for distant keys). Better sessions appear further from the origin — the near corner is beginner, the far corner is expert.
The analytics demonstrate a clear relationship: typists who focus on rhythm and using both hands see speed follow naturally. Chasing CPM directly produces erratic, plateau-prone progress. The profile makes this visible — showing a learner exactly which dimension of their technique needs attention rather than telling them to "just practice more."
Built to show a practical example: by presenting the data to his son, the underlying message — that regularity and correct two-handed technique are the levers that produce speed, not speed practice itself — became immediately tangible and motivating.
Lesson word lists are generated using LLM AI, selecting words that match the student's chosen theme — so practice material stays engaging and personally relevant rather than generic.
Beyond typing training, the app includes a spelling word practice mode: the student speaks the words they need to practise using their voice, and the tutor builds a custom spelling drill from those words — turning a school homework list or personal vocabulary goal into an interactive, repeatable exercise.
Xterra Engine Acoustic Monitor
Has anyone considered replacing some of the dozen-plus engine sensors in a modern vehicle with acoustic monitoring? Or adding a virtually free acoustic monitor to an older vehicle whose "broken sensor" check-engine light has become a permanent driving companion?
The story. A month ago my Xterra started misfiring. A week earlier the oil-pressure gauge had momentarily dropped to zero — fortunately a Jiffy Lube across the street caught a significant oil leak. The mechanic repaired the leak and refilled the oil. A week after that the misfire became noticeable. The mechanic said three ignition coils needed replacing. The question: spend on three coils, or scrap the whole vehicle?
The experiment. I aimed the same signal-processing toolkit I have spent years applying to physiological signals directly at the engine. The result is a small edge-computing web app: the mobile device records and runs light-weight processing (live RMS, spectrogram); a server runs the heavier ICA and per-cylinder decomposition; results are returned to the device.
The finding. The analysis flagged four suspect cylinders, not three.
The honest caveat. This was built fast with help from Claude Code. AI assistance is excellent for build speed and is genuinely useful for routine DSP and algorithm work. What it does not know is the set of domain-specific tricks — sensor placement, signal conditioning, isolating failure modes from operating-point drift — that decide whether a system like this actually performs in the field. Those tricks are the next steps, and they are exactly what I bring to client engagements that an AI tool alone cannot.
From Experiments to Client Method
The three prototypes above are small. Their real purpose was larger: they were the proving ground for a much faster way of running early-stage sensor-data product development.
AI-assisted prototyping changed what a real-time app costs to build. The old wall-clock for a working real-time app — once the team had worked the bugs out — was on the order of a month. With AI assistance the real-time app now comes together in hours.
That speed changes what becomes possible in week one of a client engagement. Specifically, it lets the real-time data-quality dashboard, calibration tooling, and signal-analysis pipeline — work most teams choose to defer to "phase two" — be built and used during the very first data-collection sessions. Bad data is caught while it is still being collected. Hardware issues surface in week one, not month six.
I applied this approach in a recent client engagement. The total time to a validated v1 algorithm came down significantly. The exact reduction is still being measured across more engagements, but the shape of the change is clear: risk surfaces earlier, rework loops shorten, and the team spends less time recovering from late discoveries.
Want the same methods applied to your sensor data?
The signal decomposition and ML techniques on display here are the same toolkit I bring to client engagements — in medical devices, and increasingly in aquaculture, equipment health monitoring, and edge-computing applications.
