Topic
Industry 4.0 and smart manufacturing, past the brochure
Most plants claim digital transformation while doing digitisation, and the gap between those two words is why so many programmes disappoint. Manufacturing systems, digital twins, predictive maintenance and the integration work nobody photographs.
45 pieces
Teardowns
16- Keycloak Single Sign-On Across Distributed Sites, and the Parts That Bite Realms, PKCE, token clocks, brokering and the database nobody budgeted for. What actually breaks when single sign-on spans a dozen sites, and what to decide before it does.
- Tool calling reliability: DeepSeek, Qwen, Kimi open weights and what to check before you build Four separate properties hide inside the phrase "tool calling works". Here is what each of the three Chinese open-weight families actually documents, where the calls break and how to test it yourself.
- Designing APIs for AI Agents When the Caller Cannot Read Your Docs An API read by a model has different requirements from one read by a developer. Names become prompt, errors become instructions and idempotency stops being optional once the retry loop is autonomous.
- AI Visual Inspection Fails at the Lighting Long Before the Model Classical vision still wins geometric checks. Learned models earn their place on variable and rare defects. Accuracy is set by lighting, fixturing and the false-reject threshold you can afford.
- Why a digital twin in manufacturing stalls at level two A digital twin is only useful when live telemetry feeds it and a decision comes back out. A 3D model that never updates is a rendering, and the data feed was always the expensive part.
- Where to Rent GPUs in Africa: What Is Actually Plugged In For anything larger than a fine-tune, one country on the continent has rentable modern GPUs. What is actually plugged in, what is only announced and what either costs to use.
- Six plants, one event log, and what it actually guarantees An event log across distributed plants buys ordering per partition, replayable history and one shared dependency, and most factories would be better served by a database for another five years.
- LLM Evals Are the Regression Tests for Probabilistic Software Teams that cannot measure their AI system cannot improve it, so every change becomes an argument. Evals are the regression tests of probabilistic software and the first thing to build.
- Predictive Maintenance Works. Most Programmes Still Fail. The 30 percent and 45 percent savings figures in every predictive maintenance deck trace to one uncited line in a 2010 US government guide. Here is what the primary sources actually say.
- When to Use a Service Mesh: Day One Is Usually Too Early A mesh gives you enforcement your code cannot. It also charges per pod from day one. What ambient mode changed, what a gateway plus disciplined timeouts still covers and when the tradeoff flips.
- How Many Defect Images Do You Need to Train a Defect Detection Model: 25 or 12,568 Vendors quote 100 images per class. Published work says 25 can be enough, or 12,568 may not be. The number depends on how many ways your defect can look, and here is how to measure yours.
- OT Security Breaks the IT Playbook, and AI Agents Make It Worse Availability outranks confidentiality, you cannot patch a machine mid-run, and Modbus has no password because in 1979 the wall was the password. Then we connected all of it to AI.
- Unified Namespace vs Data Historian: Why Neither Replaces the Other A UNS is a real-time topic space. A historian is an archive that discards data on purpose. The retention maths, the Sparkplug detail and the honest architecture where both exist.
- Self-Balancing Robot Simulation: What Transfers and What Does Not The inverted pendulum is the standard teaching model because it fails informatively. Simulating one is easy. Transferring the controller is not, and the reasons are always the same handful.
- Every insert writes a part Manufacturing telemetry is append-heavy, high-cardinality and almost always queried as a time-bounded aggregate. A columnar store fits that shape exactly, and breaks on updates, joins and small insert
- Generative AI for engineering design is cheap. Verification is where the money goes. The generative step in engineering costs minutes of compute. Proving the output manufacturable, simulation-clean and safe costs weeks, and that ratio decides which applications are worth doing.
Essays
11- What Happens When AI Gets Cheap, Seen From Shenzhen and Lagos Generic reasoning costs $0.05 per million tokens. The premium moved to what nobody wrote down: how one mould behaves, what a Lagos trader really charges, what a supplier means by 差不多.
- LLM API Pricing in Dollars, Revenue in Naira Token prices are set in dollars while African revenue is set in naira, and tokenizers charge some African languages nine times more tokens for the same sentence. That maths decides viability, not mode
- Made in China Quality: What the Label Actually Predicts in 2026 The label used to mean cheap and derivative. In batteries, solar, drones, robots and open-weight models it now marks the frontier. The perception lag is worth money to anyone testing on evidence.
- China Humanoid Robots Shipped 19,100 Units in Six Months. The Reducers Matter More. An ageing workforce and an explicit industrial policy point the same direction. Beneath the humanoid demos sits a component supply chain that quietly changed hands: reducers, sensors, actuators.
- AI Regulation Comparison 2026: Four Regimes, One Shipping Product Four regimes, four instruments: a product safety file in Brussels, a court docket in Washington, a filing number in Beijing and a review date in Addis. The artefacts converge. The triggers do not.
- The seniors of 2034 are not being hired this year Judgment is manufactured in the boring ninety percent of junior work. Shut the intake now and the shortage lands in the mid-2030s, when nothing can fix it quickly.
- How 3 billion yuan of bubble tea manufactured AI adoption in China China's mass AI adoption was manufactured by firms owning payments, delivery and a super-app, not by consumer curiosity. Subsidy bought usage. Whether it bought habit is a separate question.
- Shenzhen Hardware Manufacturing Is No Longer Cheap. It Is Still Four Days. Shenzhen stopped being cheap years ago. What it still sells is density: an iteration loop that runs in four days because the mould shop, the fab and the assembler are all inside an hour's drive.
- Digital Transformation in Manufacturing Is Three Different Projects Digitisation, digitalisation and digital transformation describe three different amounts of work. Most plants claim the third while doing the first, and that gap explains the disappointment.
- Sensor Fusion in Sports Tracking and What 500Hz Is Actually For A 500Hz sensor in a football is a timestamping instrument, not a measuring one. What fusion actually buys you, and why clock alignment across devices breaks more systems than sampling rate ever does.
- AI in Supply Chain: The Dull Wins and the Forecasting Trap Forecasting gets the budget and moves the least. The gains that hold up in cross-border sourcing are duller: reading documents, watching lead times drift and shrinking the exception queue.
Projects
2- Volunteer Organisation Succession Planning Starts With a Membership Record Volunteer bodies decay when the rules live in one person's head. The order that prevents it: membership record and constitution first, programmes last, with a gate between every step.
- Your sourcing agent is paid to make the order bigger Sourcing agents are paid a percentage of what you spend, so every good procurement decision costs them money. What engineer-led, fee-based buyer representation actually involves instead.
Tools
12 of 16- Instrumental review AI inspection and failure analysis for electronics assembly. Cameras capture every unit at each build state, then models flag defects no golden-sample rule ever encoded.
- ONNX Runtime review Microsoft's cross-platform inference engine for ONNX models, with execution providers for CUDA, TensorRT, OpenVINO, CoreML, NNAPI, QNN, DirectML and XNNPACK.
- Roboflow review End-to-end computer vision platform for labelling, training and deploying detection models to cloud or edge hardware like Jetson, with a visual builder for multi-step vision pipelines.
- MinerU review Document-to-Markdown engine from Shanghai AI Lab's OpenDataLab. Relicensed off AGPLv3 in April 2026, reported at 95.4 on OmniDocBench v1.6 and running on ten domestic AI chip families.
- NVIDIA Jetson and TensorRT review NVIDIA's edge inference stack: Jetson modules from Orin Nano to Thor T5000, with TensorRT compiling models into hardware-specific FP4, FP8, INT8 and INT4 engines.
- Rodin (Hyper3D) review Deemos (影眸科技), a ShanghaiTech spinout, builds Rodin. Gen-2.5 emits native quad meshes at 4k to 50k quads, which is why it survives real pipelines. Hosted only, no weights.
- Tulip review No-code platform for shop-floor apps: operator work instructions, machine and device data capture, traceability records and OEE dashboards, priced per interface with public numbers.
- Ultralytics YOLO26 review Current YOLO model family from Ultralytics: five sizes, NMS-free end-to-end detection plus segmentation, pose and oriented boxes, exportable to TensorRT, ONNX, OpenVINO and TFLite.
- Unitree UnifoLM review Unitree's open robot learning stack: world models, a vision-language-action policy, Isaac and MuJoCo training environments and the SDK that drives Go2, G1 and R1 hardware.
- BlenderProc review GPL-3.0 Blender pipeline from DLR that renders photorealistic synthetic training images with RGB, depth, normals, segmentation and exact 6D pose, exporting to BOP and COCO.
- Edge Impulse review Edge AI MLOps platform for MCU and NPU targets, covering sensor data capture, DSP blocks, training and C++ export with the EON Compiler. Qualcomm-owned since 2025.
- InternLM review Shanghai AI Lab's open-weight family (书生·浦语), now the Intern-S scientific multimodal line: 35B up to 1T parameters, Apache 2.0, trained on domestic computing infrastructure.