Blog
LLM Infrastructure
Model selection, hosting, fine-tuning, cost optimization, and scaling LLM-powered systems in production.
845 articles in this category
Affirm's Transformer Credit Model: Approvals Up, Losses Flat
Affirm says the first version of its transformer-based underwriting model added 3.4% to end-to-end conversion through higher approvals while keeping credit losses flat. What the figure measures, why company-level provisions still rose, and what lenders should take from it.
Legal AI Agents Can Call Your APIs Without Reading the Key
An ibl.ai agent can call a live external API from inside a sandbox that starts with no network, through hosts the firm allowlists, using a key it can use but never read. OpenAI, Anthropic and Microsoft ship egress controls too; the question for a law firm is who holds the key.
Germany Now Owns the Source Code of Its Combat-Drone Mission Software. What the Helsing MissionOSBw Handover Means for AI Procurement
Helsing handed Germany's procurement agency the full source code, documentation and SDK for MissionOSBw, with ownership staying with the German state. What was transferred, what is still unclear, and the procurement lesson for any agency buying AI.
Rapid-MLX 1.0: Local LLMs on a Mac, Read Honestly
Rapid-MLX 1.0 is an Apache 2.0 LLM server for Apple Silicon. What its speed numbers actually measure, what leaves the Mac by default, and where a Mac fits in an enterprise AI stack.
Ripple-65B: A 65B Apache 2.0 Coding Model That Runs on One Box. What the Benchmarks Actually Say
Ricursive's Ripple-65B Preview is a dense 65B Apache 2.0 model that reports 80.5 on Terminal-Bench 2.1, a 0.4-point edge over a 501B model. What the self-reported numbers support, what the headline leaves out, and why the license matters more than the leaderboard.
Eleven Coding Agents, One Architecture: What a Source-Code Study Means for Enterprise Teams
A Wavestone AI Lab study of eleven coding agents' source code found they converge on seven subsystems, use no vector search for code, and differ most in safety and permissions, which is where enterprise teams should spend their evaluation.
HTI-5 Dropped Model Cards for Predictive AI. The Sepsis Lesson Says Validate Locally Anyway
An independent validation found a widely used sepsis model missed 67% of sepsis cases at a hospital that had not built it. ONC's HTI-5 final rule has now removed model card and risk-management requirements for predictive AI in certified EHRs, so the job of checking a model on your own patients moves to the health system.
Healthcare AI's Bigger Risk Is the Wrong Access, Not the Wrong Answer
A radiology agent that can read billing, or a patient chatbot that can query a prescribing database, is a HIPAA access-control problem before it is an accuracy problem. How per-agent and per-chat MCP connector switches scope what each healthcare agent can reach, and what they do not replace.
Revolut Cut AI Coding Costs About 90% by Self-Hosting Open Models. What the Numbers Actually Say
Revolut says self-hosting open GLM models on 56 Nvidia B200 GPUs cut its AI coding costs to roughly a tenth of what it paid frontier labs, about $0.30 per million tokens. What the talk supports, what the viral version got wrong, and what it means for organizations without a GPU cluster.
Shannon, the Open-Source AI Pentester: What a Security Team Should Know Before Running It
Shannon, Keygraph's open-source AI pentester, passed 51,000 GitHub stars by November 2026. It is AGPL-3.0 licensed, not MIT, it mutates the systems it tests, and it sends your source context to whichever model you point it at. What that means for running offensive agents safely.
Can AI Agents Cut Platelet Waste? Hospital Logistics Evidence
Platelets last days, and a 2020 paper put US outdating near 16%. Sharing policy cut outdates about 20% at Emory, and simulation points to more. The evidence favours hospital logistics as an agent job, and the hard part is shared data.
Construct 1.0: Agents Build 3D Models and Label What They Know
Construct 1.0 turns a plain-language request into a 3D model with named parts, under Apache-2.0. It doesn't scan objects or replace CAD. Its useful idea is that every part is labelled verified, supplied, inferred or decorative.
Did Pakistan Deploy 1,500 AI Agents? What Its Sovereign AI Deal Says
The 1,500 in Pakistan's sovereign AI story is licenses for an app-building tool, not deployed agents. What the DFINITY deal covers, what the vendor now says is live, and what any public agency should ask of a sovereign AI offer.
Build, Connect, Control: The Three Layers of an Enterprise AI App
A customer service app from one prompt in about ten minutes covers only the first of three layers. Enterprise AI apps also need connections to live systems and enforceable limits on spend, access and audit, and most evaluations test only one.
Can K-12 Classrooms Use Open-Source Agent Builders Like n8n or Flowise? What Licenses and COPPA Leave Open
Open-source agent builders can replace a $49 a month Zapier plan, but none of the five we checked is built for minors. What their licenses say, what COPPA has required since April 2026, and how a district can teach agent building safely.
BCG Says Only 10 to 15% of Services Firms' AI Talent Does Applied AI. What Should Buyers Ask?
BCG found applied AI is 10 to 15% of the AI-tagged workforce at technology services firms against 90 to 95% at AI-native firms. Its conclusion is a talent decision, not a shortage, which gives buyers specific questions to ask.
Does a Government Stake in an AI Company Make AI Sovereign? What Canada's Cohere Deal Actually Buys
Canada's up to C$1.6 billion Cohere stake buys equity and company commitments, not control of any agency's AI. Sovereign AI is a property of each deployment: who holds the code, the keys, the model choice and the exit.
Which Consultancy Deploys Copilot, Gemini or Bedrock on Your Own Data? What a Deployment Partner Must Leave You Owning
Copilot, Gemini and Bedrock all promise not to train on your data, so that is not the question to ask a consultancy. The question is who owns the connectors, agents and policies it builds, and US copyright law does not default in your favor.
NemoClaw Sandboxed Agents With NVIDIA NeMo Guardrails: What Each Layer Does and How to Deploy Both
NemoClaw and NVIDIA NeMo Guardrails are two separate projects: the sandbox limits what an agent can reach, the rails check what it says. What a deployment with ibl.ai combines, and which layer handles jailbreak defense, PII, RBAC, audit and network isolation.
How to Turn a Course into AI-Narrated Video with Agentic Video
How ibl.ai's Agentic Video turns a course's scripts, slides and outlines into AI-narrated video — storyboard, presenter or generated scenes, captions and dubbing, LMS delivery — what a narrated minute costs next to studio rates, and why the files stay on your infrastructure.
Agentic LMS: The AI-Native LMS With an Agent in Every Course
Agentic LMS is ibl.ai's AI-native learning management system: an AI agent grounded in each course, AI-assisted authoring and rubric grading, on infrastructure you run, compared honestly with Canvas and Moodle.
What Is Agentic Vibe? Build Enterprise AI Apps With a Coding Agent and Deploy Them on Your Own Platform
Agentic Vibe is ibl.ai's MIT-licensed toolkit of more than 150 skills that teaches Claude Code, Codex, Cursor or Copilot to build a complete enterprise AI app, with SSO, agents, roles, memory, analytics and billing wired in, as code in your own repository.
The AMA Says Clinical Authority Is Earned. That Is an Infrastructure Bill.
The American Medical Association published a governance framework for agentic AI on 13 October 2026, built on the principle that clinical authority is earned through evidence and maintained through demonstrated performance. Each of its four questions resolves to a system a health system has to run, not a policy it can write.
Perplexity Open-Sourced Its Agent Sandbox. The Network Leaked, Not the VM.
Perplexity open-sourced SPACE on 15 October 2026 under Apache-2.0, the Firecracker microVM sandbox behind its Computer product. Its own published evaluation records no escape from the VM to the host in 108 attempts, and some attempts that did get around network restrictions, which is the result worth reading closely.
About LLM Infrastructure
Running large language models in production requires careful infrastructure planning—from model selection and hosting to fine-tuning, cost optimization, and GPU provisioning. Explore practical guides on building reliable, scalable LLM infrastructure that balances performance, cost, and latency for real-world applications.