Journal
Design choices, calculations and lessons from robotics, embedded systems, industrial automation and software. Follow a worked example, investigate a failure or look inside one of my projects.
30 posts
Rushing the first week of a project feels productive and usually costs the most time later. Why I spend longer on the problem before touching the solution, with NASA's cost data behind it.
A 0.1% difference in wheel diameter puts a differential-drive robot 0.17 m off after 10 m and 0.67 m off after 20 m. The maths behind odometry drift and the calibration that fixes most of it.
A worked ball-screw example: 20 kg moved 200 mm in 0.6 s needs 0.218 N m peak, 0.110 N m RMS and 2,667 rpm. How to get each number and which one actually decides the motor.
The same PI gains on the same saturated drive overshoot to 98% of full speed without anti-windup and settle at 90% with it. A simulated comparison of no protection, conditional integration and back-calculation.
When a retrieval-augmented system gives a wrong answer, the cause can sit in the index, the retriever, the model's reading or the final answer. How I test each stage separately so the failure points to a fix.
A September humanoid study puts mechanical design inside the learning loop. A look at its reported results, an arithmetic check, and what is actually available to reproduce.
Why a 100 kΩ NTC paired with a 100 kΩ resistor leaves the ESP32's suggested ADC range near 106 °C, and how I would size the divider for a hot process.
A practical ROS 2 fault-injection plan for stale observations, delayed commands and controlled recovery, with an illustrative stopping-distance calculation.
A detailed design study of the ESP32-S3 controller, its power and protocol boundaries, and a radar-based positioning extension that preserves the existing PCB.
Inside the SPC-001 cleaner: the built prototype, its independent research paper, and the archived CAD, traction, energy and coverage study.
How I organised sensing, estimation, planning and control for a differential-drive rover, and what simulation could establish.
Integrating Hall-effect sensing, embedded acquisition, Bluetooth and MATLAB into a movement-assessment prototype.
A local-first pricing workflow with deterministic matching, exact money handling and an operator-reviewed ServiceM8 export.
An ESP32-S3 display-controller design connecting serial parsing, guarded display output, a four-layer PCB and browser configuration.
What a product reveals about its design when someone needs to open it, find a part and put it back into service.
A small control raises a larger design question: how much attention should an everyday adjustment demand?
How an 80 Hz signal becomes indistinguishable from 20 Hz when sampled at 100 samples per second, and what to check before trusting a trend.
A worked inspection example showing why defect recall, false rejects and review workload belong beside the headline accuracy score.
Years in a role can tell us something. They cannot tell us everything. A case for assessing capability, supporting graduates and rewarding people who keep learning.
A worked example of field of view, exposure and motion blur, and how to tell whether an inspection image contains enough evidence.
What Little's Law and a simple queue model reveal about work in progress, waiting time and where automation can actually help.
Separating connection health, measurement age and validity in an MQTT dashboard, with a practical stale-data test.
A focused way to diagnose ROS 2 QoS mismatches, separate discovery from delivery, and test the actual subscriber.
A practical way to separate coordinate-frame, timing and localisation faults before changing a navigation controller.
How hysteresis, debounce and explicit states turn a fluctuating measurement into a useful event without hiding the timing cost.
A practical design for checking retrieval, source quality and permission boundaries in an assistant that runs on your own hardware.
Preserving identifiers, resolving ambiguity and making an import reviewable before it changes an operational system.
Why offline use, saved data, export and recovery need separate acceptance tests in a small personal or business app.
Designing retries, operation records and human review for the uncertain gap between sending a request and seeing its result.
A plain-English guide to trying AI, protecting your information and checking the answers.
Contact
Discuss an engineering role, a project or a research collaboration. Tell me what you are building, where it gets difficult and what a useful result would look like.
