Evaluating an LLM system before you trust it
A fixed test set, the right grader per answer type, human spot-checks and a rerun on every change, so you see what a prompt change fixed and what it broke.
Read article06 / BLOG
Things I've learned building AI.
Notes from real projects: computer vision, LLMs, datasets, deployment, working solo, and building a career around all of it.
How multi-object tracking keeps the same ID on the same person
Detection finds boxes in one frame. Tracking links them across frames. How a tracker predicts, matches by IoU, and why IDs swap when people cross.
Read articleBounding boxes vs segmentation masks: what masks actually add
A box is fast and fine for counting; a mask follows the real outline. How much background a box holds, and when masks are worth it.
Read articleHow OCR reads a shipping label, and how to catch its mistakes
Deskew, find the text, read it, parse the fields, then check them. How a check digit catches a misread before it reaches your systems.
Read articleHow RAG finds the right passage before the model answers
How retrieval-augmented generation works: chunk documents, index them by meaning, find the nearest passages, and answer with citations.
Read articleHow an AI model calls tools: the loop behind every agent
An AI agent is a model in a loop: it requests a tool, your code runs it, and the result goes back. The loop, the schema check, and the step limit.
Read articleFrom pixels to floor positions: making camera zones work in metres
A homography from four floor markers maps camera pixels to positions in metres, so zones, distances and speeds work on the real floor.
Read articleModel drift: why a vision model gets worse after launch, and how to catch it
Why vision models drift after launch, how a drift score like PSI catches it early, and how a labelling loop brings accuracy back.
Read articleDepth from two cameras: how stereo vision measures distance
Two cameras a known distance apart see the same point shifted. That shift gives depth, and explains why stereo loses precision with distance.
Read articleNon-maximum suppression: why your detector draws one box, not six
A detector proposes many overlapping boxes for every object. Here is how IoU and non-maximum suppression pick one, and what the IoU threshold does to crowded scenes.
Read articleChoosing the confidence threshold: misses versus false alarms
What precision and recall mean on a real line, and how to set a detector's confidence threshold from what each kind of mistake costs.
Read articleLighting decides what your inspection camera can see
Ring light, dark field, dome and backlight on the same part: why a scratch is invisible under one light and obvious under another.
Read articleQuantization: making a vision model smaller and faster without breaking it
What FP16, INT8 and FP8 do to a model's weights, why INT8 needs calibration, and the size, speed and accuracy trade-off you can expect on real hardware.
Read articleHow a convolution works, one stamp at a time
The operation at the heart of every vision model, explained with a 3x3 kernel, an 8x8 image, and how stride and padding set the output size.
Read articleHow a GPU runs a neural network, and why it's so much faster
Thousands of small cores, wide memory, and a batch split into tiles that run in waves. What actually happens on the card when a vision model runs.
Read articleHow an industrial camera actually captures a frame
Aperture, exposure, global shutter, readout, and the network cable: the physics behind every computer vision system, and the settings that make or break it.
Read articleWhy most AI demos never reach production
A working AI demo takes days. A system people rely on takes much longer. Here is where the gap actually is, and how to close it.
Read articleForklift and pedestrian safety with computer vision: from demo to production
How camera-based near-miss detection between forklifts and people works, why a demo is easy, and what it takes to run it on a real warehouse floor.
Read articleCounting parcels and catching jams on a sorting line with computer vision
Parcel counting, jam detection, label reading, and damaged-box checks on conveyors: what works in a demo and what it takes to run all day on a real line.
Read articleComputer vision at the loading dock: damage proof, pallet counts, and receiving checks
How cameras at the dock can count pallets, record damage, and check deliveries automatically, and what it takes to make that reliable enough for claims.
Read articleMonitoring robot palletising cells with computer vision
Using cameras to check stack quality, catch tilted crates and foreign objects, and explain line stops in robot palletising cells, from prototype to production.
Read articleVision-language models or trained detection models? Picking the right approach
When to use a VLM like Gemini, when to train a detection model like YOLO, and why most production vision systems end up using both.
Read articleTurn your backend into an agentic system: an agent layer on top of what you already run
You don't need to rebuild your software for AI agents. How an agent layer with tools, permissions, approvals, and logs sits on top of your existing systems.
Read articleWhat it took to make a simple AI website scanner production-ready
The website scanner on this site looks like a weekend project. Here's everything around the model call that made it fast, safe, cheap, and measurable.
Read articleAI agents for freight and logistics companies: quotes, tracking emails, and paperwork
Where AI agents save logistics teams real hours: quote requests, "where's my shipment" emails, and document matching, deployed safely on existing systems.
Read articleAI for online shops: product questions, custom orders, and reports that build themselves
Practical AI for e-commerce and retail brands: an assistant that knows your products, custom order quoting, weekly reports, and vision for returns and damage claims.
Read articleYou don't need to be a tech company to have a developer
Any business can now have one experienced developer build its internal tools and automations, replacing subscriptions for less than a full-time hire.
Read articleAI didn't make everyone a 10x engineer
AI coding tools amplify experience instead of replacing it. Why a lot of do-it-yourself AI builds stall at the demo, and what actually makes the difference.
Read articleWhy I build solo (and why it's often faster than a team)
AI coding agents changed what one developer can do. Why one person who builds everything often beats a team, or even one hire, on speed, quality, and cost.
Read articleDesigning browser agents that fail safely
A practical workflow for browser automation agents with scoped tools, checkpoints, evidence, and recovery paths.
Read articleHow to design a production computer vision pipeline
A practical architecture for taking computer vision from a notebook to an observable, maintainable production system.
Read articleA practical method for optimizing computer vision inference
How to find the real bottleneck across decoding, preprocessing, GPU inference, tracking, and result delivery.
Read articleBuilding reliable LLM workflows beyond the demo
How to combine models, tools, retrieval, validation, and observability into LLM integrations that teams can trust.
Read articleBuild an AI engineering portfolio that proves the work
A project structure that shows technical depth, product judgment, deployment experience, and clear communication.
Read articleValidate your vision dataset before training
The checks that catch expensive annotation and distribution problems before they become model problems.
Read articleBiology is becoming a software system
A practical way to think about the convergence of AI, biological measurement, intervention, and programmable medicine.
Read articleFive years of learning in public changed my career
How publishing simple technical tutorials created skills, opportunities, a global network, and a long-term personal brand.
Read articleSoftware engineering is the other 90 percent of production AI
Why useful AI systems require architecture, interfaces, deployment, scaling, and operations around the model.
Read articleWhy computer vision remains a strong AI specialty
Real-time visual systems still require dedicated models, data, software architecture, hardware, and domain engineering.
Read articleReal-time dense 3D reconstruction with MASt3R-SLAM
What makes monocular dense reconstruction valuable, how the pipeline works, and what engineers should validate before deployment.
Read article