Topics

AI in production

Name: senn-tech
Address: Weissachstraße 67, Kufstein, Tirol, 6330, AT
Telephone: +4369917602100
Price range: €€

RAG, agents and automation in the company: which setups hold up in operation, what needs maintenance, and where the models stop working.

58 posts

2026-10-24 · KINemotron 3.5 Embed, Rerank, Parse: Nvidia's open RAG models from 1BAt GTC Berlin Nvidia presented three open models for retrieval pipelines: Embed (1B and 8B), Rerank (8B) and Parse (0.63B). Every benchmark figure comes from the vendor. What this means for a self-hosted RAG system.
2026-10-24 · KIKimi K3.1 Pro: 4th in the AA Index, open weights announced, 2.8 trillion parametersMoonshot AI launched Kimi K3.1 Pro on 23 October: 4th place in the Artificial Analysis Intelligence Index, 2.8 trillion parameters, weights promised by 4 November. We keep measurements, vendor claims and our own arithmetic apart.
2026-10-24 · SoftwareKeePassXC 2.8: Wayland Auto-Type, database sync, Qt 6KeePassXC 2.8.0 has been out since 21 October. Auto-Type now works on Wayland through XDG portals, the Qt 6 migration is done, and the new Polkit quick-access feature changes the database file. What to check before you update.
2026-10-24 · KIStudy: the agent harness drives token costs up to 3x, the pass rate barely movesFive models, three harnesses, one benchmark: Liu and Han measure that with the model held fixed, the token bill changes a lot and the success rate hardly at all. What that means for running coding agents yourself, and how to measure it.
2026-10-19 · Local AIOpen-weight models: 71.5% of tokens and 23% of spend on the Vercel AI GatewayVercel's gateway data shows open models handling over 70 percent of tokens, with their share of spend rising from about 5 to 23 percent in three months. Useful numbers for the open-versus-closed-API argument, with clear limits.
2026-10-16 · KI & EntwicklungTorvalds has Claude Code rework the page-cache read path: what the numbers show and what they do notLinus Torvalds published eight experimental patches in a branch which Claude Code wrote under his direction. On his laptop the kernel reads the same file up to three times as often per second. What was measured, what is missing and what an operator takes from it.
2026-10-16 · AI & DevelopmentOpen AI models you can self-host in October 2026: what fits in 24, 128 and 512 gigabytes and what the licence allowsThree weeks brought several open models that a business of 20 to 100 people could run in its own building: Nemotron 3.5 Bolt and Super VL, LFM2.5, MiniCPM-V 4.7, EmbeddingGemma 2, two lean Qwen variants and Meta Muse Spark Open. We read the model cards, the licence texts, the file sizes and the Hacker News threads, and sort the models by the memory they need.
2026-10-14 · AI & DevelopmentCloudroom Measured: An Open Kernel, One Author, Four Weeks OldA founder presents his cloud agent platform: a sandbox for every agent, an open kernel, a free plan. We measured the repo, the licence, the pricing and the security page; the video's superlatives do not survive the sources.
2026-10-14 · AI & DevelopmentAnimation as code: what Claude Motion does and what a Remotion licence costsClaude Motion builds explainer clips from program code instead of a video model. We read the documentation, worked through Remotion's licence tiers and rebuilt the mechanism on our own hardware.
2026-10-12 · AI & DevelopmentFork-and-Flush: AI agents work better when you wipe the chat historyAn arXiv paper fixes the stalling of long agent runs by keeping the work and discarding the chat. Several branches run in parallel. Numbers, limits and what you can take from it for your own agents.
2026-10-11 · AI & DevelopmentA Website with AI in Eight Phases: What a Video Promises and What We MeasuredA free German skill walks Claude Code through building a whole website: brand, keyword research, copy, design, launch. We checked the video's numbers against GitHub, the README and our own archive.
2026-10-11 · AI & DevelopmentStrata runs a 125-billion model at 50 to 60 tokens per second on a graphics card from 2020A self-test is making the rounds: the RX 6800 that crawled at 3 to 4 tokens per second under Ollama in August now writes 50 to 60 on the Strata backend. We checked the math: part of the headline does not hold.
2026-10-11 · AI & DevelopmentREA hit 83,000 GitHub stars in 44 hours. We measured the curveReverse Engineer Anything is a real, MIT-licensed MCP tool for taking other people's software apart. Its popularity is measurably manufactured. Here is the curve, the method behind it, and the two traps we walked into.
2026-10-11 · AI & Development84 days to decompile a game, twelve clone apps in a week: what software actually costsA Nintendo 64 game fully decompiled in 84 days, twelve open-source programs modelled on Adobe and Microsoft products inside one week. We measured both stories against the repositories and started all twelve programs in a browser.
2026-10-11 · AI & DevelopmentAnthropic takes AI tests offline: agent sandboxes need egress rulesAnthropic has switched off live internet for all internal model evaluations after models submitted real forms, exploited a university server and pulled access tokens. The cases, Anthropic's measures and what follows for your own agents.
2026-10-07 · AI & DevelopmentMistral Large 4 among the open weights: index scores, prices and the bytes behind themMistral Large 4 is supposed to ship open weights at the end of October. We checked the independent index score, the prices and the checkpoint sizes of the Chinese models, then measured our own inference lane against them.
2026-10-07 · AI & DevelopmentClaude Haiku 5.5: $0.10 per million tokens, but heavy token useWith Haiku 5.5 Anthropic cuts small-model prices by roughly 75 percent. Artificial Analysis measures about three times the token use of GPT-6 Luna. What that means for classification and subagents.
2026-10-05 · AI & DevelopmentAleph Alpha's Kolibri: a German tokenizer that will not fit our GPUsKolibri landed on Hugging Face under Apache 2.0 with 78 billion parameters and a tokenizer built for German. We measured the tokenizer against our own production lane and priced the hardware.
2026-10-05 · AI & DevelopmentClaude Mods: What the New Plugin Format in Claude Code Is Allowed to DoClaude Code 2.1.287 shipped mods on 1 October and they are on by default. They run as code inside the process, with the user's permissions. As of 11 October: ten releases in eight days, a catalogue of over 3,000 community mods, and what managed settings actually block.
2026-10-04 · AI & DevelopmentLocal AI in a Field Test: Seven Days Without Cloud Models on 64 GBAdrian Twarog spent a week running nothing but local language models on a 64 GB mini PC. The models held up; the context budget inside the coding harnesses did not. Here is the field test next to our own production setup.
2026-10-04 · AI & DevelopmentParakeet Redux: 178 MB of speech recognition, one proprietary kernelMoondream ternarised NVIDIA's Parakeet down to 178 MB and improved it across 25 languages. We checked what survives on our own hardware and read the engine's licence text.
2026-10-04 · AI & DevelopmentClaude drives FreeCAD over MCP: what the CAD agent builds and what it must not touchSince 15 July 2026 a video by the Make Form channel shows Claude Desktop modelling real parts in FreeCAD over the Model Context Protocol: boxes, fillets, even a spur gear with no gear workbench installed. Behind it sits the community project freecad-mcp, 2,654 stars, MIT licence, an XML-RPC bridge on port 9875. This post sorts what the demo actually proves, what the project docs say about security that the video leaves out, and where the line sits for real engineering work today.
2026-10-03 · SoftwareExchange Online has been blocking EWS since 1 October 2026, and Apple Mail on the Mac now depends on the AppID allow listSince 1 October 2026 Microsoft has been flipping the Exchange Online setting EWSEnabled from Null to False, tenant by tenant, which blocks Exchange Web Services. On 1 April 2027 EWS is switched off for good, with no exceptions. Apple Mail, Calendar and Contacts on the Mac still speak only EWS; Apple has announced Graph support for a future macOS 27 update without giving a date.
2026-10-01 · AI & DevelopmentAutomating an ERP with AI: what actually attaches to a WinLine systemOn our WinLine ERP the AI path came down to access levels: one database user with SELECT rights only, seven read tools for an agent, and exactly one write tool.
2026-09-26 · AI & DevelopmentLocal AI is not only getting better, it is also getting smallerFour releases, one direction: llama.cpp v0.5.0 and vLLM v0.30.0 change operations, ternary weights and MiMo v2.6 move the size floor. One of our boxes has been running the new vLLM since the weekend.
2026-09-22 · AI & DevelopmentJev in Practice: A Model That Delivers Decisions, Not TextJev is not a chatbot but a decision model: three question types, milliseconds, fractions of a cent. We checked the model against published run costs, guardrail patterns, and the open alternative Laya.
2026-09-22 · AI & DevelopmentBonsai 2 27B: A 27B Model at 1.7 Bits, and the Fork It ForcesPrismML compressed the exact base model of our production lane, Qwen3.8-27B; into a ternary GGUF at 1.72 bits per weight. Five days in: 2.2 million downloads, a fork requirement, and benchmarks we have not reproduced.
2026-09-11 · AI & DevelopmentEUR 261 instead of EUR 7,640 in API fees, a month of agent traffic, measuredA LiteLLM gateway, vLLM on four RTX 5090s, an agent walking the company over SSH. Tokens, cache hits and real costs, measured, not projected.
2026-09-07 · AI & DevelopmentGPT-6 Astra: 99.9 percent on ARC-AGI-3, and why the benchmark foundation still will not say AGIOpenAI has released GPT-6 Astra, and president Greg Brockman declared the AGI era. The same week delivered the counter-calculation: 62.7 instead of 99.9 percent in the standard harness, an Intelligence Index level with the predecessor, and a model that is harder to monitor than the one before it.
2026-09-05 · AI & DevelopmentQwen3.8-27B with Claude Opus reasoning traces: four fine-tunes tested, one holds upAn active community trains open models on Claude Opus reasoning traces. We checked four candidates for our production model Qwen3.8-27B, rebuilt one ourselves, and ran it against our own test battery.
2026-09-04 · AI & DevelopmentA rented GPU instead of our own production: testing vLLM's offload path for Qwen3.8-Flash-NextTo find out whether vLLM can offload Qwen3.8-Flash-Next's huge embedding table to host RAM, we didn't risk our own production box; we rented a matching GPU machine for two hours instead, and needed three failed attempts before the first working launch.
2026-09-02 · AI & DevelopmentvLLM Prefix Caching in Self-Hosting: Measurably Fewer TokensSince 19 August 2026, our local Qwen3.8-27B not only returns answers but also reports which prompt tokens were reused from the KV cache. This is how we measure the real effect of prefix caching.
2026-08-30 · AI & DevelopmentQwen3.8-27B Self-Hosted: Running Locally on 4× RTX 5090Since 18 August 2026, our local stack runs Qwen3.8-27B-FP8 on four RTX 5090s, replacing the previous Qwen3.5-122B-A10B MoE. Why a smaller dense model replaced the giant, and what changed in daily operations.
2026-08-29 · KI & EntwicklungFreeToken: Frontier MoE Models on Gaming HardwareFreeToken is a new edge-native MoE serving engine that runs 290B+ frontier models on consumer hardware using one GPU plus CPU/RAM.
2026-08-28 · KI & EntwicklungKTransformers vs FreeToken: Two Paths to Big MoE Models on Small GPUsTwo frameworks promise to run huge MoE models on a single GPU. Where KTransformers and FreeToken aim at the same target and where they differ.
2026-08-27 · KI & EntwicklungSelf-Hosted LiteLLM Gateway: Two Lanes Instead of a WildcardOur local LiteLLM gateway was consolidated on 19 August: only thinking and nothink on the 4× RTX 5090 host, hermes on the single-card box.
2026-08-18 · AI & DevelopmentSpaceX has bought Cursor: who owns the developers' tools nowAs of 14 August 2026 Cursor is a SpaceX subsidiary, 60 billion dollars, entirely in stock. What the largest startup acquisition ever means for companies that use the tool daily.
2026-08-09 · AI & DevelopmentOpen models running on proprietary hardware versus closed cloud APIsIn-house GPU or third-party API? The answer depends on three factors: quality gap, utilization, and the value of your data. We use both and break it down here.
2026-08-01 · Tools & ProductivityHandy: The Free, Offline Speech-to-Text Tool That Actually WorksHandy turns speech into text, completely offline, free, and open source. We tested it, gathered community feedback, and explain why it deserves a spot in your toolchain.
2026-07-26 · AI & DevelopmentWhen AI Runs Your Infrastructure: Three Lessons for a Production-Grade MCP ServerLetting a chat agent clone VMs, take snapshots and migrate containers is magic in a demo; it only becomes production-ready through permission profiles, semantic tool routing and self-healing connections. Three lessons from the field.
2026-07-24 · AI & DevelopmentOmniRoute vs. LiteLLM: Two Self-Hosted AI Gateways, Two Different JobsTwo self-hosted, MIT-licensed LLM gateways, but built for different jobs. LiteLLM manages AI usage across a company; OmniRoute is the local-first power tool for one developer. We run both, and here is which belongs where.
2026-07-22 · AI & DevelopmentCoding Tools for Windows: Claude Code, OpenCode, Codex, Cursor, and Orca Put to the TestA Comparison of Five AI Coding Tools on a Windows ARM64 Computer: Which ones run natively, which ones require workarounds, and which tool is best suited for which workflow?
2026-07-21 · AI & DevelopmentOrca: The IDE Built to Run a Fleet of Coding Agents in ParallelOrca from Stably AI is an open-source Agent Development Environment: run Claude Code, Codex, and 30-plus CLI agents side by side, each in its own isolated git worktree. We use it daily; here is why the worktree-per-agent model changes how you work.
2026-07-06 · AI & DevelopmentOne RTX 5090 Instead of Four Old GPUs: Consolidating Vision InferenceFor the image and PDF stage of our extraction pipeline, we replaced four older accelerators with a single RTX 5090. Why one current card beats four old ones here, and what the real root cause was.
2026-06-23 · AI & DevelopmentMaking Self-Hosted AI Faster: Speculative Decoding and dFlashSelf-hosted inference is decode-bound, not compute-bound. Speculative decoding fixes that without touching output quality, and NVIDIA's new open dFlash drafter pushes it further. What it is, the numbers, and what it means for our own GPUs.
2026-03-20 · AI & DevelopmentCodex, Gemini CLI & Co: The Agent Landscape 2026Claude Code, OpenAI Codex, Google Gemini CLI; the most important agentic coding tools at a glance. Strengths, differences, and why we take a multi-track approach.
2026-02-12 · AI & DevelopmentClaude Code in Practical UseAnthropic's agentic coding assistant in real project life: What we use Claude Code for, how Skills, MCP servers, and subagents work together, and where the limits lie.
2026-01-15 · AI & DevelopmentAgentic Coding: How AI Agents Transform the Dev Daily RoutineFrom autocomplete to autonomous agent: What 'agentic coding' really means, where it works today and where human judgment remains indispensable.
2026-01-01 · AI & DevelopmentEmail AI: Classification, Routing, and Reply Suggestions LocallyAutomatically classify emails, route to the right handler, and generate reply suggestions, with local LLM, n8n, and own mail infrastructure. Productive, not futuristic.
2025-11-06 · AI & DevelopmentMCP-Server in-house: Agents with their own knowledge fedModel Context Protocol gives coding agents tools, but the most valuable MCP servers are those with their own knowledge: internal docs, ERP data, TMS status. So we build them ourselves.
2025-09-16 · AI & DevelopmentEU AI Act: Why Local AI Is the Right AnswerThe EU AI Act regulates AI applications by risk class, and makes locally operated models the strategically cleanest option. What the law requires and why owning your hardware simplifies compliance.
2025-08-19 · AI & DevelopmentAI Phone Agents: Voice Bots with Pipecat & AsteriskReal-time voice pipeline, STT, LLM, TTS; connected to a phone system: Pipecat and Asterisk make it possible. A look at what works today and where the limits lie.
2025-07-15 · AI & DevelopmentWorkflow Automation with n8n, Self-Hostedn8n connects services, webhooks, and now also AI agents in a self-hosted workflow tool. What sets n8n apart from Zapier and Make, and why we use it for serious automation.
2025-06-17 · AI & DevelopmentRAG Environments Compared: Vector Databases, Chunking, EmbeddingsRetrieval-Augmented Generation promises facts from your own documents, but the pipeline behind it decides the quality. Vector databases, chunking strategies, and embedding models in a practical check.
2025-06-03 · AI & DevelopmentLLM Frontends: OpenWebUI vs. LibreChatTwo self-hosted web interfaces that turn local LLMs into a usable chat service. OpenWebUI and LibreChat compared: multi-model, RAG, user management, everyday usability.
2025-05-20 · AI & Developmentllama.cpp & GGUF: LLMs locally, efficiently, without a clusterllama.cpp enables LLM inference on consumer hardware; quantization, CPU/GPU offload, and a file format that needs no cloud budget. The counterpoint to vLLM and when to use which tool.
2025-05-06 · AI & DevelopmentHigh-Throughput LLM Inference with vLLM on Custom HardwarevLLM makes LLM inference fast and efficient, PagedAttention, continuous batching, and an OpenAI-compatible API. Here's how we run it on an RTX 5090 and what really matters.
2024-12-10 · InnovationHow to Embed AI in Your Company Without Losing ControlArtificial intelligence has become a competitive factor. How to embed AI securely and sustainably in your company, from data foundation to operation.

AI in production

Other topics