Blog
A 2026 technical archive across AIML, computer vision, 3D CV, NLP, LLM, VLM, VLA, VLNA, embodied AI, medical robotics, MLOps, IoT, robotics, quadcopter design, Fusion 360, Autodesk workflows, special effects and software delivery.
Featured
We moved a production logistics dashboard from React 18 to React 20 across three sprints, with the client's traffic running the whole time. Cold builds went from 141s to 58s, the bundle shrank 22%, and two of our starting assumptions turned out to be wrong — the full metrics and the diffs that mattered are in here.
Read the articleAll posts
Showing 1–6 of 113 posts · Page 1 of 19
Quantization is not a checkbox at export time — it is an engineering pipeline that spans training, calibration, compilation and on-device validation.
6 min read
Read the articleRegulators have opened their first AI Act cases over lab incidents that happened before release, turning agent containment from an engineering nicety into a compliance question.
5 min read
Read the articleNew research removes the internal ‘instruction’ signal from untrusted tokens, cutting injection success to near zero on a standard benchmark, but it only works where you control the weights.
6 min read
Read the articleAn 816-billion-parameter open-weight model that activates only 42 billion parameters per request, and fits on a pair of desktop workstations, shows how sparse architectures are changing the economics of self-hosted AI.
5 min read
Read the articleShipping a new model to thousands of edge devices is a distributed-systems problem with an ML-shaped blast radius; here is how to build the release path so a bad model is caught on ten devices instead of ten thousand.
9 min read
Read the articleThe language model is not a security boundary, so guardrails for production agents have to be enforced in code around it: scoped capabilities, provenance tracking, sandboxes and approvals bound to exact actions.
8 min read
Read the articleAn 816-billion-parameter open-weight model that activates only 42 billion parameters per request, and fits on a pair of desktop workstations, shows how sparse architectures are changing the economics of self-hosted AI.
5 min read
Read the articleA major bank is adding a thousand contractors to fix flaws surfaced by frontier AI models, and hundreds of scientists want cyber-capable models air-gapped by default. The bottleneck in software security has moved from discovery to remediation.
5 min read
Read the articleWhen splitting work across AI agents actually pays off, and how to engineer the coordination layer like the distributed system it is.
7 min read
Read the articleHow a trained graph becomes fast, correct code on edge silicon, and where each stage of the compiler pipeline can quietly change your model.
8 min read
Read the articleWhy production agents need an explicit execution graph, and how checkpointing, idempotent side effects and trajectory replay make them resumable, auditable and testable.
8 min read
Read the articleHow to split inference between a small on-device model and a larger upstream model so you get near big-model quality without saturating the uplink.
9 min read
Read the articleWhy agent memory is a data-engineering problem, and the tiered design we use to keep long-running agents grounded, affordable and auditable.
9 min read
Read the articleHow we take a float model to integer kernels on edge NPUs and GPUs without losing the accuracy that justified building it.
8 min read
Read the articleMost agent failures in production happen at the tool boundary; here is how we make tool calls validated, effectively-once and recoverable.
7 min read
Read the articleHow to size, build and benchmark an edge GPU pipeline that sustains many camera streams at a known latency and accuracy, not just a fast single-camera demo.
8 min read
Read the articleInference is moving off the cloud and onto cameras, gateways and microcontrollers — here is what it takes to ship it reliably.
4 min read
Read the articleOn-device language model inference is a memory-bandwidth problem long before it is a compute problem; this is how we size, quantize and accelerate it without losing the behaviour we tested.
8 min read
Read the articleFinal-answer accuracy hides most of what goes wrong inside an agent; this is how we build evaluation harnesses that score the whole trajectory, survive non-determinism and gate releases.
8 min read
Read the articleFrom ant colonies to drone fleets, swarm intelligence shows how decentralised agents following simple rules can produce remarkably smart collective behaviour.
3 min read
Read the articleLanguage models learned from text; embodied AI learns from acting in the physical world — and that changes everything about how we build it.
3 min read
Read the articleBeyond simple chat prompts, autonomous AI agents are executing complex, multi-step workflows across enterprise software.
2 min read
Read the articleA practical comparison of model choice tradeoffs for teams balancing quality, response time and monthly runtime cost.
7 min read
Read the articleA retail shelf-audit workflow moved from manual photo checks to a review queue that flagged missing facings and price-label mismatches.
7 min read
Read the articleWhat to confirm before moving vision inference from a cloud prototype to an on-device production workflow.
6 min read
Read the articleA practical note on using depth maps and point clouds to inspect dents, missing parts and alignment issues.
6 min read
Read the articleHow intent clusters reveal which support queues are ready for automation and which need cleaner process definitions.
6 min read
Read the articleA grounded look at device gateways, MQTT topics, retry queues, dashboards and alert rules for industrial monitoring.
8 min read
Read the articleA procurement assistant stopped guessing part numbers once retrieval, null answers and approval gates were treated as product features.
8 min read
Read the articleHow depth cameras, LiDAR and calibration routines shape reliable perception pipelines before autonomy is added.
7 min read
Read the articleUsing VLMs to sort field photos into defects, context shots and unclear evidence before a technician reviews them.
6 min read
Read the articleWhat changes when a robot policy reads a visual scene, follows language instructions and still needs strict action limits.
7 min read
Read the articleA design note on matching payload mass, battery capacity, motor selection and realistic endurance targets.
6 min read
Read the articleA facility team used vision-language-navigation assistance to guide inspections through confusing service corridors.
7 min read
Read the articleConstraint setup, load cases, manufacturing limits and material selection for practical generative design studies.
5 min read
Read the articleWhy simulated agents need messy sensor logs, interruptions and recovery states before real-world trials.
8 min read
Read the articleA non-diagnostic robotics workflow for handoffs, checklists and equipment readiness in controlled healthcare environments.
6 min read
Read the articleVersion control, comments, review packages and handoff discipline when mechanical teams are not in one room.
6 min read
Read the articleModel releases need eval gates, rollback plans and cost checks just as much as code releases need tests.
7 min read
Read the articleRetrieval structure, chunk quality, metadata and evaluation methods for internal technical knowledge bases.
8 min read
Read the articleA short launch film became feasible after segmentation masks, review passes and compositing handoff were made repeatable.
6 min read
Read the articleHow lighting, camera angle and false-positive review changed more than the model architecture.
6 min read
Read the articleHow sampling frequency, radio usage, sleep modes and enclosure constraints affect device life in the field.
6 min read
Read the articleA field team replaced repeated site visits with a 3D reconstruction workflow and a structured exception list.
7 min read
Read the articleA clause extractor works best when uncertainty is routed clearly instead of hidden behind confidence scores.
6 min read
Read the articlePackage structure, simulation loops, launch files and hardware interfaces for teams moving beyond lab demos.
7 min read
Read the articleA repeatable prompt review process covering golden sets, regressions, retries and user-facing failure messages.
7 min read
Read the articleA selection framework covering sensor support, tuning workflow, logs, failsafe behavior and maintainability.
5 min read
Read the articleTechnicians stopped resending unclear photos once the VLM flagged missing angles before submission.
6 min read
Read the articleWhy VLA systems need hard limits, simulation checks and explicit recovery states before touching physical workflows.
6 min read
Read the articleEvent schemas, retention windows, idempotency and query patterns for products that ingest continuous device data.
7 min read
Read the articleHow maps, landmarks, camera pose and spoken instructions combine into usable navigation support.
7 min read
Read the articleWe moved a production logistics dashboard from React 18 to React 20 across three sprints, with the client’s traffic running the whole time. Cold builds went from 141s to 58s, the bundle shrank 22%, and two of our starting assumptions turned out to be wrong — the full metrics and the diffs that mattered are in here.
9 min read
Read the articleHow to define inspection metrics, sample sets and human review loops before training the first production model.
8 min read
Read the articleA warehouse pilot improved faster once operator corrections were captured as training data instead of complaints.
8 min read
Read the articleA safe operations-focused view of tracking instrument presence, handoff state and audit records.
6 min read
Read the articleA practical decision guide for device-to-cloud communication, buffering and operational visibility.
6 min read
Read the articleWhat changes when the monitored input is an image, a caption, a sensor stream and a generated answer.
7 min read
Read the articleMaps, dynamic obstacles, aisle width, recovery behavior and the operational limits that decide path planning quality.
7 min read
Read the articleA production note on mesh cleanup, texture review and delivery constraints for interactive 3D demos.
6 min read
Read the articleA calibration-first guide for teams trying to measure objects across several camera viewpoints.
7 min read
Read the articleParametric modeling, cloud review, simulation and CAM handoff patterns that shorten prototype cycles.
6 min read
Read the articleForty-plus systems, six months of logs: where the models saved real hours, where they quietly created work, and the three we switched off.
11 min read
Read the articleFiltering noise, aligning scans and turning messy point clouds into useful engineering evidence.
6 min read
Read the articleLayouts, data refresh patterns, permissions and incident views for operators and engineering teams.
7 min read
Read the articleHow language detection, translated queries and source citations make multilingual retrieval less fragile.
6 min read
Read the articleDuplicates, labeling drift, class imbalance and train-test leakage checks that save weeks of model iteration.
6 min read
Read the articleSupport tickets reached the right queue faster after the model was constrained to routing, not answering everything.
7 min read
Read the articleUsing VLMs to compare annotations, callouts and part references before engineering review.
6 min read
Read the articleHow to match measurement accuracy, environmental rating and calibration effort to the actual operational question.
5 min read
Read the articleA practical checklist for simulation coverage, blocked actions, unsafe states and human approvals.
7 min read
Read the articleSimulation, recorded sensor bags, hardware-in-the-loop checks and regression tests for robotics teams.
8 min read
Read the articleWhy the best navigation dataset may be the log of every time the assistant got confused.
6 min read
Read the articleThe retrieval setup we reach for when the source material is 400 PDFs, two wikis and a shared drive — and why chunking is the part that decides everything.
8 min read
Read the articleWhere each sensor helps, where it fails, and how fusion improves stability and navigation confidence.
7 min read
Read the articleHow to collect synchronized video, commands, actions and outcomes without burying the team in unusable data.
7 min read
Read the articleUsing simulation outputs to focus design reviews on stress, tolerance, thermal and manufacturability risks.
6 min read
Read the articleA therapy-support dashboard helped staff compare sessions, spot missed exercises and keep review notes consistent.
7 min read
Read the articleA guide to separating live operational views, diagnostic views and business summaries without duplicating products.
6 min read
Read the articleWhy many first production models need disciplined snapshots before they need a feature store.
6 min read
Read the articleModel registry, rollback plans, monitoring, dataset snapshots and release gates for teams crossing into production.
8 min read
Read the articleCombining depth, tracking and compositing previews so creative reviews happen before final render time.
6 min read
Read the articleHosting, model calls, monitoring and the occasional human review — an itemised monthly bill from a real deployment, not a pricing-page estimate.
6 min read
Read the articleWhere digital twins help facility teams, and how to avoid modeling details that never affect decisions.
6 min read
Read the articleA mixed OCR and image-quality workflow for reducing label mistakes before goods move downstream.
6 min read
Read the articlePractical design choices for override controls, handoff states and recovery when autonomy needs help.
7 min read
Read the articleHow to evaluate depth outputs against the robot task instead of generic leaderboard scores.
7 min read
Read the articleRoute planning, geofencing, battery reserves and abort behavior for safer inspection workflows.
6 min read
Read the articleMatching duplicate customer records with text similarity, rules and a final human merge step.
6 min read
Read the articleModel setup, toolpath review, fixtures and tolerance communication before parts leave the screen.
5 min read
Read the articleA checklist for security, latency, evals and maintenance before choosing a private model deployment.
7 min read
Read the articleA practical note on where agents should suggest, where they should execute, and where humans should approve.
6 min read
Read the articleHow visual models can prepare review notes while keeping safety decisions with qualified staff.
6 min read
Read the articleCredential storage, firmware updates, network segmentation and logging checks for connected prototypes.
7 min read
Read the articleReorder points, demand variance and a forecast that only had to beat a spreadsheet — the working behind the number on our home page.
7 min read
Read the articleA robot cell became reliable after point-cloud cleanup, grasp logging and operator feedback were connected.
8 min read
Read the articlePayload, reach, repeatability, guarding and end-effector decisions for compact automation projects.
6 min read
Read the articleLighting, occlusion, privacy and workflow limits that shape safe computer vision around sterile environments.
6 min read
Read the articleA design walkthrough for balancing thrust, current draw, heat and endurance in quadcopter builds.
5 min read
Read the articleContent models, drafts, roles, slugs, previews and migration planning for replacing the first editorial release.
6 min read
Read the articleRoughly one prospect in five leaves the first call with a no-AI recommendation. Here is the checklist that gets them there.
5 min read
Read the articleThe inspection team caught a camera-angle drift before quality dropped because image stats and review outcomes were monitored together.
7 min read
Read the articleWhere synthetic samples can improve coverage, and where they hide labeling or lighting problems.
7 min read
Read the articleWhere synthetic scenes help visual pipelines, and where they create artifacts that reviewers immediately notice.
6 min read
Read the articleHow common sensor streams become maintenance signals, and what baselines are needed before alerts mean anything.
8 min read
Read the articleTurning field failures, operator notes and repair patterns into better hardware and software decisions.
5 min read
Read the articleHow inspection footage becomes tagged findings, reviewed evidence and assigned maintenance work.
6 min read
Read the articleSketch discipline, named parameters, assemblies and versioning habits that keep designs editable.
6 min read
Read the articleThe one-page structure we use to fix scope, price and the definition of done before any code exists. Copy it freely.
4 min read
Read the articleLogs, traces, quality checks and alert rules that show whether automated workflows are helping or drifting.
7 min read
Read the articleHow to show accuracy, failure examples, drift and human review outcomes in a way operators can use.
6 min read
Read the articleAlert grouping, status hierarchy, trend windows and device health signals for practical operations screens.
5 min read
Read the articleMilestones, risk registers, prototype scope and acceptance criteria for technical projects with hardware or models.
7 min read
Read the articleNo posts match that search.
Filed under
Newsletter
Every post above landed in the newsletter first. Practical AI and automation notes from our projects — no hype, no forwarding your address, unsubscribe anytime.
Sent Thursdays. Around 600 words, usually with a number in it.
Prefer the live version? Book a demo