Diepgaande artikelen van het Software Tailor-team over compliance voor gereguleerde AI, on-prem-uitrolpatronen, onze producten en de marktontwikkelingen die ze vormgeven. Elke bewering is voorzien van bronnen; elk artikel is gedateerd en wordt bijgewerkt wanneer de onderliggende feiten veranderen.
De AI Omnibus trad op 27 juli 2026 in werking en verplaatste de hoogrisicoverplichtingen naar 2 december 2027. De verplichtingen veranderden niet, alleen de datum — en 16 maanden zijn goed voor één architectuurwijziging, niet twee.
The sticky routing strategy maps each caller to a preferred worker by an identity hash, so multi-turn chats keep landing on the box where the model is loaded and the prompt cache is hot — a deterministic mapping with no session table to replicate.
Mark a worker canary and give the pool a percentage: that share of callers, assigned consistently by identity hash rather than per-request coin flips, prefer the canary while everyone else stays on stable — and the two groups remain each other's failover.
Per-worker concurrency limits keep each GPU box in its sweet spot. A saturated worker is skipped; a fully saturated pool returns 503 with Retry-After immediately — honest backpressure instead of a silently growing queue that ends in a thrash or an OOM.
A served AI Server deployment's security posture in plain statements: fail-closed serving, keys hashed at rest, two credential tiers in a farm, forwarded pseudonymous identity, fingerprint pinning, and private TLS to workers — the default behavior a security review reads.
Every AI Server hosts a live dashboard and native Prometheus metrics, and its content-free audit attributes each request to app, install, and serving worker — so per-team, per-app chargeback is native data, no metering sidecar required.
A gateway that discovers its workers over mDNS on the local network, so scaling a small site is plugging in a prepared box — it advertises itself, the gateway probes it, and within a health-check cycle it is serving. Static core and dynamic discovery compose.
Serve both generations at once during a migration: modern OpenAI-compatible clients on the governed primary port, and pre-v2 clients on a scoped compatibility listener that carries a compiled sunset date — no flag-day, no broken scripts.
The compliance-hardened deployment: region-aware governance that fails closed where content scanning is mandated, on-device moderation, per-key quotas and budgets, and a content-free, signed audit an external auditor can verify offline, independently.
Upgrade workers and the gateway while traffic keeps flowing. A draining /readyz probe, a configurable drain hold, and gateway failover cover the requests racing the drain — no maintenance window.
Run one AI Server on the office GPU box and serve every workstation on the LAN over an API-key-authenticated endpoint. The model loads once, the data never leaves the building, and there are no per-token bills.
Run AI Server as one container on a cloud VM or any Docker host: configured entirely by environment variables, with /livez and /readyz probes and JSON logs so orchestrators manage it correctly. The image is the artifact; the volume is the state.
Every AI Suite app creates one loopback AI Server on your machine. It binds 127.0.0.1 only, needs no keys or licence, and shares one model in memory across every app on the box — the zero-exposure default worth understanding before anything is networked.
A farm whose workers carry different model sets. The gateway routes each request to a box that already has the model warm, and advertises the union of every worker's models as one catalog — so callers see one model list and placement is the gateway's problem.
One URL and one key in front of a pool of worker AI Servers: health-checked load balancing, transparent failover, model-aware routing, and Kubernetes auto-discovery — with no vendor cloud in the request path.
A ground-up rebuild of the shared engine behind AI Suite: host, download and run the AI models every app uses — one download, one library, one set of keys, used everywhere. Free serves this machine; Pro serves your whole network with governance. Free on Windows from the Microsoft Store.
Every Software Tailor AI app — Chatbot, Translate, Rewrite, Notepad, PDF Reader, Image Generation and more — in one window, with the local AI engine built in. Each app keeps its own model and history; your work stays on your device, saved encrypted. Free on Windows from the Microsoft Store; one Pro subscription unlocks the whole suite.
The full private AI chat for Windows: multi-session chats, folders, a prompt library, conversation branching, document chat and durable memory — all free on a local model that works offline, with encrypted on-device history. Pro adds tool use, Agent mode and larger or cloud models. Your existing chats and prompts upgrade in place. Free from the Microsoft Store.
Type a prompt and an original picture appears on your own PC — a built-in FLUX.2 [klein] model, negative prompts, seeds, batch generation and a private encrypted gallery. Free at 512×512; Pro unlocks larger sizes, custom steps and optional bring-your-own-key cloud. Free on Windows from the Microsoft Store.
A photo assistant that turns a folder of pictures into a searchable collection — AI captions, object detection, OCR, opt-in face grouping and semantic search, all on your device. Organise, annotate and remove backgrounds. Free on Windows from the Microsoft Store.
A ground-up rebuild: pick a model, type, and a local AI model streams the reply back on your own device — with an encrypted local history, find and search, and two simple context controls. Deliberately slim. Free on Windows from the Microsoft Store.
A ground-up rebuild built for reasoning models that think step by step — maths, logic, code — with a collapsible Thinking panel showing the chain-of-thought and a larger reply budget so answers aren't cut off. Free DeepSeek R1 model. Free on Windows from the Microsoft Store.
Type any text and hear it spoken in a natural voice — and clone a voice of your own from a short clip — right on your device. Offline by default, encrypted at rest, with a seekable player and WAV/MP3 export. Free on Windows from the Microsoft Store.
A simple notepad that formats itself as you write — plain text to Word, automatically — with on-device AI writing help, ask-your-document Q&A with citations, ask-across-all-open-documents, annotations and a readable dark editor. Free on Windows from the Microsoft Store.
Turn speech into text on your device — from your microphone, your computer's audio, or an audio file. Export subtitles, replay and re-transcribe recordings, keep an encrypted local history. Free on Windows from the Microsoft Store.
A ground-up rebuild: paste or import text, choose an action, tone and length, and a local AI model rewrites it on your device — with match-my-style, document import and an encrypted local history. Free on Windows from the Microsoft Store.
A ground-up, cross-platform rebuild: open a PDF and ask questions answered on-device, with clickable page-exact citations — in a full reader with annotations and bookmarks that travel in the file. Windows, macOS, Linux, plus a headless CLI/MCP server.
De architectuur van enterprise AI heeft in 24 maanden de defaults verschoven. EU AI Act-verplichtingen voor deployers, de NIST AI RMF-richting voor kritieke infrastructuur en a16z-dealcyclusgegevens wijzen allemaal dezelfde kant op: residency-first is het nieuwe cloud-first.
Software Tailor levert al 19 jaar maatwerksoftware zonder projectmislukking. Vier engineeringregels houden die staat overeind. Het verhaal van de oprichter over de discipline achter het cijfer.
De OECD AI Policy Observatory volgt AI-beleid in meer dan 80 jurisdicties. De AI Policy Toolkit van juni 2026, de GPAI–OECD-fusie en het omnibusakkoord rond de EU AI Act zijn de drie signalen die er voor enterprise-inkoopteams toe doen.
AI Admin Console biedt zes capaciteiten — Members, Licenses, AI Server-enrolment, Policy, Audit, gebruik per organisatie. Elk beantwoordt een specifieke vraag die inkoop- en compliance-teams stellen onder de deployer-verplichtingen van de EU AI Act en het NIST AI RMF.
Elke beheeractie in AI Suite landt in een JSONL-auditregel met tijdstempel, actor en actie — en geen prompt- of responsinhoud. De deployer bewaart de inhoud; wij bewaren het bewijs dat een actie heeft plaatsgevonden.
Enterprise AI-inkoop loopt vast op verplichtingen van de deployer, niet op technologie. De eigenschappen van de uitrolarchitectuur die de impasse in 2026 doorbreken, en waarom "we hebben SOC 2" het verkeerde antwoord is op de vraag die compliance-teams nu stellen.
De verplichtingen van de Act voor deployers van hoog-risico AI worden gefaseerd ingevoerd in 2026. On-prem-uitrol ontwijkt de regelgeving niet, maar verandert wel welke verplichtingen praktisch haalbaar zijn, en welke door cloud-SaaS in het geheel niet vervuld kunnen worden.
Lokale inferentie, audit onder controle van de klant en een inkoopbeweging die al bestaat. Drie structurele redenen waarom Software Tailor's AI-producten desktopinstallaties zijn en geen tenant in onze cloud.
Er worden maandelijks meer artikelen toegevoegd. De volledige catalogus, inclusief de backlog van nog te schrijven onderwerpen, wordt bijgehouden in onze repository.
Als er een regelgeving, uitrolpatroon of branchevraag is waarover u wilt dat wij schrijven, laat het ons weten. Wij kiezen elke maand één nieuw artikel.