The archive
Everything published, newest first. Four disciplines, one continuing argument: that good systems are built by people who understand why, not only how.
Two questions every articulated mechanism asks: given the joint angles, where is the tip? Given a tip pose, which angles get you there? The math is the same whether the arm is on a desk, a factory floor, or a simulation.
Most enterprise AI programmes fail for the same reason enterprise search failed in 2008: nobody owned the data. Here is what actually has to be true before a model earns a place in production.
Looking back across thirty years of systems, the choices I am still glad about have almost nothing in common technically — but they share a single quality: they kept the next decision cheap.
Published in Clinical Rheumatology — what four years of annual biosimilar rituximab retreatment showed for disease control and safety in seropositive rheumatoid arthritis, and why real-world evidence still matters beside the trial literature.
From the GPU in your cloud instance to the quantum chip that may one day break encryption — a map of 19 processor architectures shaping 2026 and beyond, with practical guidance on when each one actually matters.
The strangler pattern is well documented and still routinely misapplied. The hard part was never splitting the code — it was splitting the data while both halves stayed correct.
A conference paper from ICON BITS 2023 (BITS Pilani), published by Springer — on low-cost, community-level STEM pathways, age-appropriate toolkits, and why access is an engineering problem as much as an education one.
Twenty-seven methods, one question behind each: does this pattern mean something, or could it just be noise? Grouped by what they are for, with the reasoning and a worked example behind every one.
The best debuggers I have worked with were not the ones who knew the codebase best. They were the ones willing to be proven wrong quickly, in public, on purpose.
How a network turns a wrong guess into a better one — one derivative at a time. The chain rule, activation derivatives, the softmax–cross-entropy shortcut, and a worked example you can trace by hand.
NodeMCU ESP8266 + MPU6050 gyro/accelerometer with SSD1306 OLED readout, optional DHT11, InvenSense teapot packets over serial, and a Processing 3D visualiser — motion on the bench and on the screen.
One equation from 1961 predicts why your service falls over at 80% utilisation and not at 100%. It is the most useful mathematics I have ever applied to a production system.
Raspberry Pi Hass.io as the hub, Mosquitto MQTT, NodeMCU driving an 8-channel relay bank, Home Assistant entities, an optional NodeMCU WebSocket UI, and a lighter IR/mobile alternative without the Pi.
Wiring a 1-inch hall-effect water flow sensor to a Particle Photon — pin map, Photon features, pulse-to-L/min math, and firmware that publishes instantaneous and cumulative flow to the Particle Cloud.
A DIY 3018 CNC mill driven by GRBL and CNCjs on a Raspberry Pi — milling wood and soft metal from CAD designs, laser engraving, and remote job control from any browser on the home network, built for under $200.
Turning an old Raspberry Pi into an Echo Dot–like Alexa Voice Service device — USB mic and speaker, Amazon developer profile, AVS sample app, companion service, Java client auth, and Sensory wake word — for home automation and voice Q&A.
Assembling a Tevo Tarantula DIY kit — full parts list, commissioning steps from bed levelling to Repetier Host G-code, and the first successful print — roughly $200 and 8 kg of desktop manufacturing.
Building a first quadcopter from a 330 mm frame — brushless motors, ESCs, KK2.1.5 flight controller, RF gear, camera pan/tilt, mechanical and electrical assembly, calibration, and PID notes — about ₹10K and 1 kg, tested indoors only.
A Waveshare Alphabot-based home robot on Raspberry Pi + Arduino — remote web/mobile drive, pan-tilt camera with MJPEG streaming, line tracking, IR remote, obstacle avoidance, and Axelta sensors for climate, light, reed and PIR.