Operations & AI insights

Insights for running a business that doesn’t need you.

Practical operations and AI insights to help your business become easier to run, and less dependent on you.

Title card: Testing An AI Agent Before It Goes Live, First 100 Days Operations

How Do You Test An AI Agent Before It Touches A Live Business Workflow?

Test an agent with an eval set of 20-50 tasks from real work, sandbox the tools, run it in shadow mode beside your current process, graduate autonomy on evidence, and turn every failure into a permanent regression test.

31 October 20269 min read

Technical asset

Your AI Agent Doesn't Have a Memory. It Has a Context Window -- And That Is the Problem

Agent memory fails quietly: context degrades answers and one poisoned write persists. What to store, what to expire, and how to test the store.

30 October 20269 min read

Title card: a 1.3B parser reads messy scans while a separate 2B checker model decides which extracted pages need human review

Technical asset

A 1.3B Model Reads Your Scans. The Interesting Part Is the 2B Model Checking Its Work

OpenBMB released a 1.3B document parser alongside a separate 2B checker that scores OCR output and flags what a human must review, and the split is what makes document automation safe near financial records.

29 October 20265 min read

Title card on a light ground with an orange rule, reading DeepSeek open-sourced the engine behind its AI service, with the line The margin is in the cache, not the model.

Method

DeepSeek Open-Sourced the Engine Behind Its AI Service. The Interesting Part Is What It Kept

DeepSeek open-sourced DeepInfer, the engine behind its own AI service, and disclosed what it earns per GPU-hour. The parts that make those economics work stayed inside the company.

27 October 20265 min read

75 per cent price rise figure beside the post title on a light title card

Proof asset

Local AI or Seat Subscriptions: The Pricing Just Moved on Both Sides

OpenAI added $100 Premium seats and NVIDIA's DGX Spark rose 75 per cent in a year. What each route costs a 25-seat firm, and the register worth building before you commit either way.

26 October 20267 min read

Technical asset

Lovable Moved Its Own Site Off Next.js. The 98.5% Figure Is the Story.

Lovable has moved its own 42-million-visitor website onto the same server-rendered hosting as its 70 million AI-built apps. The migration is engineering news; the 98.5% lift in traffic from AI tools after pre-rendering is the number owners should look at.

25 October 20268 min read

Title card: Your Agent Does Not Need a Memory Layer. It Needs Its History Back.

Technical asset

Your Agent Doesn't Need a Memory Layer. It Needs Its History Back.

A plain retrieval baseline over raw conversation history beat the tested agentic memory systems in six of seven benchmark settings, and took a fraction of the ingestion time. Measure the cheap baseline first.

24 October 20266 min read

Technical asset

One Box, Five Models, and a Licence Problem

LocalRouter runs chat, speech, image and video on one NVIDIA DGX Spark with a sub-second voice loop. The router is Apache 2.0. The models are not, and the licences decide what a business can ship.

23 October 20268 min read

Technical asset

The expensive model should stop typing: what the handoff research actually found

Escalating a cheap model's failed work up to a strong one recovers less than half the quality gap and multiplies cost. The favourable direction is the other way round: strong model plans, cheap model executes.

22 October 20267 min read

What gets written here

Five pillars, one question each

Method: How do data consolidation and task-specific AI agents work in an operating business?
Adviser asset: Which operational issues create diligence risk or owner dependency?
Proof asset: What data and system evidence will an acquirer need to trust?
Outcome asset: How can a founder-led business become more transferable?
Technical asset: How can AI be used safely in core workflows?