Notes from the handwriting frontier.
Plain-language write-ups of the research that shapes how Vahini works, how machines reconstruct strokes from motion, why two views of a letter beat one, and what your handwriting can reveal about your health. Every post links to the original paper.
All posts
Read by topic.
How Vahini reads handwriting: from a photo to a 20-factor report
A plain-language tour of how Vahini turns one photo of handwriting into twenty quality scores and a coaching plan, finding the writing, measuring the factors, reading the words, and building the report.
How AI is learning to spot learning disabilities early
Neural networks now flag dyslexia, dysgraphia and dyspraxia from reading patterns, handwriting and motor coordination. Here is how the models work, and where Vahini's motion signal fits.
How Vahini can support autistic students who write
For many autistic students, handwriting carries hidden motor and sensory load. A dual-IMU pen that measures pressure, rhythm and effort can make that load visible, and support gentler, calmer practice.
The startup pitch deck guide (with a free template)
What actually goes into a deck that gets a meeting, the 11 slides investors expect, how we framed Vahini's, and a free downloadable pitch-deck guideline plus a sample Startup India Seed Fund application.
One IMU or two? Why dual-IMU sensing is more accurate
A single inertial unit can't tell a tilt from a slide. Putting two IMUs in one pen, at different points along the barrel, resolves the ambiguity and sharpens every downstream factor.
Reading pressure from an analog force sensor
A force sensor gives a single wavering voltage. Turning it into a clean, calibrated pressure signal, and learning what heavy and light pressure say about control, is real signal work.
Reading pen tilt from an IMU
Gravity is always pointing down. By reading how it splits across an IMU's axes, and fusing gyroscope data to fight drift, the pen knows exactly how it's being held.
Velocity: the speed signal inside every stroke
Fluent writing has a velocity fingerprint, fast on the straights, slowing into the curves. How the pen recovers speed from acceleration, and why it matters for complex Indic letters.
Motion: turning movement into strokes
Motion is the master signal the others ride on. How the pen reconstructs the path of the tip from raw inertial data, and renders the conjuncts of Indic scripts faithfully.
Rhythm: the music of fluent handwriting
Skilled writing is rhythmic, a steady beat of strokes and pauses. How the pen measures that tempo, and why rhythm is one of the clearest signs of a confident hand.
Intent: knowing when the pen is really writing
Half of a pen's motion isn't writing at all. Detecting the writer's intent, pen down to write, pen up to move, is what keeps the dotted i's and stacked Indic vowel signs honest.
The magnetometer: direction, zeroing and a cleaner signal
An accelerometer and gyroscope drift in heading over time. A magnetometer supplies an absolute compass reference that zeroes the system and steadies orientation, vital for script direction.
Streaming sensors over BLE: packets, fragmentation and flow
Bluetooth Low Energy was built for sips of data, not a firehose of motion samples. How we packetise IMU streams, beat MTU limits and fragmentation, and keep every sample in order.
Reconstructing handwriting from motion alone
A pen only touches the paper part of the time. Teaching a model to tell writing from hovering, and rebuild the trajectory from raw inertial signals, is the heart of motion-based handwriting.
Writing without a screen: surface-free pens
Classical sensor fusion drifts within seconds. How learned models reconstruct multi-stroke writing from an inertial pen on ordinary paper, no tablet, no special surface.
104, 208 or 416? Choosing how often a pen should look at your hand
Our sensors can sample at 104, 208 or 416 times a second. We ship 208. The physics of handwriting bandwidth, what doubling the rate really buys, and what it silently costs.
Two views of a letter: the image and the strokes
The finished image of a letter and the time-ordered strokes that drew it are two different views. New work feeds both to a single model, and reads handwriting better than either alone.
Handwriting as an early health signal
From childhood dysgraphia to the micrographia of Parkinson's, the way we form letters carries motor-health information, and machine learning is learning to read it.
The roadmap is less hardware, not more
Most gadgets add parts each generation. Our roadmap points the other way: fewer components, a simpler board, a cheaper pen, with accuracy as the one thing that is never traded. How software absorbs hardware.
The hardest part: reading a hand it has never seen
A model trained on one set of writers and devices stumbles on the next. “Domain adaptation” is how researchers teach handwriting AI to generalise to a brand-new hand.
Why the pen's AI gets better every month it exists
Handwriting models are not finished artifacts; they compound. More writers, more ages, more scripts, more edge cases, each loop through the data makes every user's pen sharper. On the flywheel and its honest limits.
Rebuilding the pen tip's path: the hardest easy-sounding problem we work on
Turn motion into the exact path the tip drew. Simple to say; the physics fights back with drift, gravity and geometry. The three enemies of reconstruction and the structural choices that beat them.
Taming the compass: magnetometer noise cancellation in practice
The pen's compass would be a superb orientation reference, if desks weren't made of steel and laptops weren't magnets. How calibration and cross-checking turn a jittery compass into a trustworthy one.
The sensor that thinks: event engines inside the IMU
Modern motion sensors carry tiny programmable state machines that detect taps, drops and pick-ups on their own, while the processor sleeps. How the pen delegates its reflexes to silicon.
The force sensor: why the pen needs to feel the paper
Pressure scoring is the obvious reason the tip has a force sensor. The deeper reason: it is the anchor that keeps motion-based reconstruction honest. On pen-down truth, drift resets, and pressure as a signal.
Chasing microamps: low-power engineering inside the pen
Milliamps decide whether the pen lasts a day. Microamps decide whether it lasts a month in a drawer. Inside the sleep-state ladder, wake-on-motion, and the discipline of accounting for every microamp.
Bluetooth and Bluetooth Low Energy are not the same thing
They share a name, a logo and a radio band, and almost nothing else. Why the pen speaks Bluetooth Low Energy, what actually differs under the hood, and what that buys you in battery life.
Engineering a battery that writes all day and survives its owner
A pen battery must last a school day, charge over lunch, shrug off being left on the charger all weekend, and protect itself from an owner who forgets it exists. How we engineer for all four.
The most important part we didn't design: the standard refill
Building around a standard ballpoint refill sounds unglamorous. It is one of the best product decisions we made. On ink, geometry, supply chains, and respecting a hundred years of pen ergonomics.
There is a whole computer inside this pen
Processor, memory, radio, sensors, power management and a filesystem's worth of storage, all inside a barrel you hold like any other pen. A tour of the Vahini pen as the small computer it really is.
No posts under that topic yet.
The science is moving fast. So is your handwriting.
See your own 20-factor breakdown, free, then come back to read how it works under the hood.