JEV - quick test: Customer feedback analysis
Spend some time to play with JEV. Definitely there is way more to exolore
Posts & launches
What I shipped, what broke, and what I would do differently. Short pieces, written as things happen.
Slow risk and late alarms: reading the cash before it harms
A full borrower-risk system read from one balance snapshot: metrics, weights, alerts, rollout and the feedback loop that keeps it honest.
Keeping part of the thinking physical
Voice is becoming the main interface, which is exactly why a pen, paper or a stylus still earns its place in the loop.
The setup I work from when I am not at a desk
The tablet is the screen, the phone handles input and voice, and a paper notepad catches whatever will not wait.
A library of communication patterns, not one tone of voice
The same intros and the same replies keep getting rewritten. Better to keep the patterns and assemble the right one.
If I install this, what am I not doing?
A new tool always looks like progress. The question that kills most of them is what it displaces.
Still one of the best productivity tools on the market
No subscription, no battery, no login, no regional restrictions, and no limit on what shape a thought can take.
A weekend with an old Raspberry Pi and an older speaker
Bluetooth to a Bose Link, a small interface for Radio Paradise, and AirPlay on top. Found parts, one weekend.
Credentials stay behind an explicit approval step
The policy sits in the app, not in a prompt or a habit. Locking and unlocking by hand buys real peace of mind.
The calendar as an operations layer for agents
It already tells me what happens when. The same timeline can schedule an agent and record what it actually did.
Giving an agent a card, a limit and a ledger
An agent that stops at the payment step loses most of its value. Virtual cards, budgets and approvals fix that.
Speaking about Huffman coding at the School of Troubleshooters
A talk on the algorithm and the discussion around it. Nerve-racking, and worth doing again.
An assistant is a core plus the capabilities you plug in
The useful question stops being what the assistant can do and becomes which capability to add next.
Always go to the gemba
Ten minutes behind a counter teaches more than another hour of analytics. Selling it yourself teaches more again.
Heartbeat, the skill that wakes up and looks for work
Every few hours it checks activity, calendar and goals, then asks what it can move forward without being told.
A private clipboard for passing data on site
Client networks rarely make data exchange easy. A password protected clipboard on my own domain does.
An agent deleted the document and nothing could recover it
Backups turned one bad action into an annoyance. The lesson was bigger: guardrails are not bureaucracy.
Rest is also work
Most of my notes get written while walking. Past a certain point, staring at the screen returns very little.
Brief gives direction, debrief gives memory
One page in the morning on what deserves attention, a few lines at night on what actually changed.
Company research as a repeatable script
The same questions in the same order, then a lens on top. That is the difference between a dossier and a move.
A last checkpoint before anything becomes public
The problem is never a lack of care. It is volume. A scanner on outgoing files catches what you stopped checking.
Nexoder, a terminal interface for the servers I keep running
Health checks, key stats, services and the ones I use most, stored locally. Built after looking up the same commands too often.
Decisions should belong to the system, not to whoever was in the room
One decisions.md in the project root: what was decided, why, what else was considered, what it changes, who it affects.
Step away from the screens
Walking, air, and pines in the wind. It connects more dots than another hour in front of a monitor.
An agent that knows it is blocked, and who to ask
One blocked agent is manageable by hand. Dozens are not: escalation has to live inside the system.
Forgotten access is someone's opportunity
People change roles and leave, permissions stay. A live map of who has what, and why, makes zero trust operational.
Reports moved from Markdown to HTML
Markdown is good for generating and poor for reading. One template, one colour per project, one folder to open.
Feed the AI the distillate, not the landfill
Scripts collect and clean the data, sort the signal by topic and sentiment, and every conclusion stays traceable.
Keep the work where the work happens
Issues become tasks, pull requests become threads, and context sits next to the change. Agents work better for it.
One agent flagged a prompt injection left by another
It found an instruction telling it not to tell me, refused to continue, and showed it. Even your own agents need watching.