Free
AI Visibility & GEO in Practice
Four-engine measurement, the five rings, and a real case: 0 to 25% AI mention rate in six months.
INSIGHTS
What we think, and why it changes what you should build.
Decision-support writing, not thought leadership. Each piece exists because the same question keeps coming up in real conversations: what category is this, why did our RAG pilot stall, should we license or build, and how do we know an AI-generated video is broken before we ship it. Each gives an answer you can act on rather than a position you can agree with.
4
decision pillars
6
engines feeding the knowledge
5
product lines as the field
3
languages in parallel
Whitepapers
Free
The Cheap Model Costs the Most
A 53× spread on one prompt, three max_tokens traps, silent cache failures — all measured in production.
Start here
Field Notes
AI Video Generation: The Hard Part Isn't Generating It — It's Knowing It's Broken
Run AI video as a business and the bottleneck isn't model quality — it's whether you can tell a clip is broken before you ship it. Today's generative models can…
CATEGORY DEFINITION
What is an AI Knowledge Infrastructure Company?
An AI Knowledge Infrastructure Company turns an organisation’s verified operating knowledge into reusable technology platforms — not into slide decks, and not i…
ENTERPRISE RAG
Enterprise RAG that does not hallucinate — and stays maintainable
Most enterprise RAG pilots do not fail at retrieval. They fail three months later, when the source documents have moved, nobody owns the index, and the answers …
DECISION GUIDE
License, build, or consult?
There are three ways to get an enterprise AI platform, and the right answer depends on one question more than any other: do you intend to own this capability in…
From the engineering side
Production notes on enterprise AI architecture, knowledge operations, and the systems that keep them reliable — written from what we actually ran, with the numbers left in.
PVL.AI Tech Notes · AI Radar
107 Teams Patching the Same Hole: Your AI Forgets You
Of 2,282 community plugins, 120 solve one thing — memory that survives across sessions — from 107 unrelated teams. The official directory has one.
PVL.AI Tech Notes · AI Radar
Two Thousand Plugins. Who Checked Them?
291 official and 2,282 community plugins from 1,764 upstreams. The directory does verify — pinning, identity, static checks. It just does not block.
PVL.AI Tech Notes · AI Radar
What Did You Just Hand Over?
The MCP/A2A protocol war ended in 2026. But authorization is optional in the spec, and 108,245 servers grow by 519 a day. Speed, not safety.
PVL.AI Tech Notes
Four Counterintuitive Lessons From Months of LLM Experiments: The Cheap Model Is the Expensive One
Four lessons from production: route different tasks to different models, cheap models often cost more in total, paying for a top model to get it right once beats repeated retries, and a max_tokens set too low turns your entire spend into nothing.
PVL.AI Tech Notes · Tech Blog
Two AIs, Opposite Answers, the Same Mistake — Why a Locale Semantic Baseline Is the Real Moat
Measured: same model, same pipeline — error rates differed by an order of magnitude before and after the glossary was in place. Two different model families made the same class of error when no baseline existed. The ceiling on AI throughput is baseline coverage, not model capability.
PVL.AI Tech Notes · Tech Blog
Enterprise RAG Is Harder Than the Model: Designing a Multi-Tenant Architecture
A practical guide to enterprise RAG multi-tenancy: PostgreSQL RLS, connection-pool safety, pgvector, hybrid retrieval, embedding migrations, and operational controls.
Have the question these do not answer?
These cover the questions we hear most. If yours is different, tell us what you are trying to decide — the answer is usually shorter than a page.