[LOG] /blog — 26 entries
Building with AI, in production.
Notes from the work itself — orchestrating multiple models, automating the busywork nobody wants, and making AI clean up after its own mistakes. Written for people who buy this stuff, not just people who build it.
Reproduction-Driven Development and Our Multi-Model Eval Setup
How we use Reproduction-Driven Development (RDD) combined with GLM-5.3 and Flash models in a continuous evaluation harness to keep autonomous coding agents honest.
The Agent You Installed Still Doesn't Know Your Business
Pre-built agent-persona libraries like agency-agents put 200-plus specialists a click away. The costume is generic; the part that actually protects your business is the exceptions you still have to write yourself.
Do You Need a Control Plane for Your AI Agents Yet?
A real tool called LoopX is the hook for a bigger question: does your business have multi-session, multi-agent work yet, or is this infrastructure you don't need.
Is Automating This Making You Worse At It?
A real concern taken seriously: automating a task can erode the judgment its repetition used to build. Here's the test for which tasks are actually at risk.
One API Key for Everything — What That Convenience Actually Costs You
API aggregators promise one key for every service. That consolidation removes real friction, but it also removes the separation that once kept one provider's bad day from taking down the rest.
Real-Time AI Video Editing Is Arriving — What It's Actually For
A research project edits video frame by frame from a plain-language instruction. Here's what that concretely unlocks for a small business, not a production studio.
Should You Self-Host the Tools You Depend On?
1Panel makes self-hosting easy, and the monthly cost comparison usually favors it. The real question is who owns uptime, patching, and backups once the bill disappears.
The Free API Is Not the Expensive Part
Picking a free API is the easy part. The real cost shows up later: breaking changes, rate limits at scale, and quiet shutdowns you never planned for.
Turning a Free API Into a Product Isn't a Business Model, It's a Starting Point
Wrapping a free API in a paywall is a weekend project. Whether anyone keeps paying you instead of building it themselves is a different question entirely.
What a Competitive-Intelligence API Actually Tells You
A price and review API gives you a fast, structured read on competitors. It cannot show you their margin, their inventory, or why a customer picked them instead.
What Happens the Day Claude Is Down
Once a Claude workflow becomes how the task actually gets done, an outage stops being hypothetical. A practical, non-alarmist way to size the plan to what's really at stake.
When Scraping a Competitor's Site Is Smart, and When It's a Liability
Modern scraping tools make competitor price monitoring trivial to build. The real question was never technical — it's what robots.txt and terms of service permit, and what happens when the site changes.
You Don't Need Every Agent Tool on the List
Curated AI tool lists are great for discovery. Evaluating every entry before picking one costs more time than any single tool saves.
Automating Your Workflows With Claude: Where to Start
Not a chatbot habit but a workflow habit: how to pick the right repeating task, give Claude context once, and know where a human still needs to check the output.
Give Claude the Context Once: Building a Reusable Project for a Repeat Task
A practical setup guide: what instructions, reference files, and examples go into a reusable Claude project so you stop re-explaining the same task every week.
One Workflow, One Job
Why folding a third task into an existing Claude setup makes all of them worse, and the checkable signal that tells you it's time to split it into two.
When to Let Claude Just Do It (and When to Make It Wait for You)
A per-task test for autonomy: weigh the cost of a wrong output reaching someone outside your business against how easily you'd catch it, and know the gate can tighten again.
65,000 Part Numbers, and Almost No AI
Loading a supplier catalogue into Magento is the data-entry job everyone wants to throw an LLM at. Here is what actually did the work across 16 catalogues and 102,000 rows — and the one place the AI earned its keep.
NO ESCAPES FOUND
A security review came back clean, confident and well-structured. It would have satisfied my commit gate. It had never run anything at all.
The Qwen Pro Plan Mistake
The Qwen Pro plan was supposed to protect my Claude budget. The model impressed me; the allowance didn't hold up, and that gap is the whole story.
You Are Paying Salary Rates to Type
Nobody has data entry in their job title, so nobody has ever costed it. Here is what a week of typing actually costs, the three questions that decide whether it is worth automating, and the jobs where it is not.
ClawWatch: How I Force AI to Fix Its Own Code Errors (Automatically)
A PostToolUse hook for Claude Code that catches its own mistakes with Ruff and Pytest, then forces a fix loop before the agent is allowed to call anything done.
Meet Umisaiah: The Next-Generation AI Expert for Motorcycle Apparel & Accessories
Generic AI chatbots hallucinate safety ratings and sizing. Umisaiah is a purpose-built shopping assistant for powersports gear that separates expert knowledge from live inventory and refuses to guess.
Automating the Paddock: How AI Powers My Weekends
Zero time for busywork when you're spending half the summer trackside — why the AI stack behind this blog exists to clone the founder and buy back weekends.
3 Quick-Win Automations You Can Deploy This Week (That Actually Move the Needle)
Automation doesn't require a six-month IT overhaul. Three high-impact automations — lead capture, email triage, invoice parsing — you can ship in days.
Beyond the Chatbot: How I Orchestrate My Multi-Model AI Stack
The future of AI isn't one omnipotent chatbot — it's a fleet of specialized models, each doing exactly what it does best. Here's the current lineup and what's actually being built with it.
NEXT.RUN
Have an idea worth automating?
I build the same kind of systems this blog writes about — multi-model pipelines, automated data entry, and tools that fix their own mistakes. Tell me what's eating your team's time and I'll tell you if AI can take it off your plate.
Bring Your Idea to Life →