Hey, I'm Daxesh π
I build AI-powered SaaS products from scratch.
What I've built:
β Wooffer: 600K+ users
β 12 AI agents that run for @ctrlops_ai
β Products touching 2.5M+ lives
What I post about:
β How to build with AI (practical, not hype)
β DevOps simplified
β Founder lessons from 0 β scaling
I started coding on a βΉ1,000 government tablet.
Follow if you're building something π
At @blrtechsummit today and you look after Linux servers? Come to stall 4C181.
We're showing @ctrlops_ai there: deploy, debug and manage Linux servers with AI. Your SSH keys stay encrypted on your own machine.
See it running and talk to us, the people who build it.
Today is the last day of the summit, and the last day of our 31% off summit coupon on the lifetime plan.
Here in Gujarat, businesses worship their new account books on Diwali night.
Some now put a laptop in the puja, beside the books or in their place. Many still use paper.
Happy Diwali.
What would you put in front of the gods: your laptop, or the production server?
Recently, a reputed institution emailed us. They wanted to buy.
Before buying, Swaminarayan Gurukul, Rajkot compared @ctrlops_ai with three other tools, including foreign ones. They chose ours, an AI terminal for DevOps and Linux servers, and bought 5 licenses.
Honestly, we didn't see this coming.
It wasn't about brand or origin. Everything they needed was in one place: file management, Git operations, one-click deployment, AI terminals. No switching between tools.
When buyers compare tools, the one that holds the whole job beats the bigger name.
Thank you, Swaminarayan Gurukul.
When you compare dev tools, what decides it for you: features, price, or how many tabs it saves you?
The hardest part of building with AI: deleting the tool you were proud of.
I built 10 agents that turned an SRS and a DFD into development prompts. The PM agent read the SRS, spotted what was missing and asked about it before the build, not halfway through.
I tested it on one project, on my own machine. Then the models caught up and I shelved it.
No regret. Drop the tool. Keep the pattern behind it.
Is there a tool on your team that only exists because an older model couldn't do the job?
My planning agent forgets everything every week. On purpose.
It starts each week in a new chat. Nothing carries over.
Everything that matters lives in files: the plan, one file per task, a playbook of lessons. Before it plans anything, it reads them.
The other agents save their work in files too. They send the planner only a short summary and the file path, so its chat stays small.
A chat is a scratchpad. Files are the memory. A fresh start can't carry last week's confusion, and anything worth keeping gets written down where every agent can read it.
It's the split I build every agent on: the rules that stay the same go into the agent, what changes goes into a playbook.
Where does your agent's memory live: in the chat, or in files?
I took a full day off a couple of weeks ago for a close friend's wedding. No work at all.
The hard part was the week before it.
The Saturday before was already an off day. So the posts for the day after the wedding had to be ready and approved before the weekend started. On the wedding day itself, posting just paused. Planned, not missed.
Consistency is my weak spot. What I'm learning: it isn't willpower every day.
It's planning the days you can't show up, before they arrive. Then a day off doesn't break the chain.
What do you set up before a day off so nothing breaks?
Your AI agents won't tell you they're late. They'll just be late.
Everything my agents make has one deadline rule: ready for my review the working day before it's due.
But a rule nobody checks is a wish. So here's the setup:
1. Write the deadline as a rule, not a hope.
2. Give ONE agent the job of checking every open task against it, every morning and every evening.
3. Send the alarm where the late agent has to read it. Say exactly what's late and by when.
4. The checker never does the late work. It only raises the flag.
Mine is my Personal Assistant agent. Its alarm is a file in the planner agent's inbox, and a new file there wakes the planner at once.
It already caught next week's work not being ready, days before it was due. The planner filled the gap the same day.
The rule is the same for every task, so it lives in the system, not in each agent's instructions.
Who checks your agents' deadlines today: another agent, a script, or you?
Sources:
GitHub AI Scan for pull requests (GA): https://t.co/4iT7uL0bp3
GitHub agentic autofix (GA): https://t.co/rfyympyGYd
Claude Code Review (GA): https://t.co/f7nsCDkC9Y
The Register on Bank of America's 1,000 contractors: https://t.co/ZO7dW13W3A
Switching on an AI code reviewer is now a checkbox. Dealing with what it finds is not.
On 15 Oct, GitHub's AI Scan for pull requests, GitHub's agentic autofix and Anthropic's Claude Code Review all went GA.
Per The Register, Bank of America's CEO said the bank brought in 1,000 contractors to patch flaws an AI bug-finder had found.
Finding issues got cheap. The human queue behind them didn't.
I run a team of AI agents. They produce work all week, and nothing goes out until I approve it. Output is cheap. My review is the scarce part, so the system protects it: work is ready a day early, and I review in a fixed slot.
AI review is worth turning on. Just give the queue an owner first, and decide how much of it gets fixed each week. Count findings fixed, not findings found.
If you've switched one on: who owns the findings on your team?
Lifetime plan or monthly? Every founder selling a dev tool argues about this one.
One day last week I got a small, real data point.
CtrlOps, our AI terminal for DevOps and Linux servers, had 3 sales in a single day.
2 picked the lifetime plan.
1 picked monthly.
Three buyers is not a pattern. And I don't know why each one picked what they did.
Still, it's a real signal from people who actually paid, not from a poll.
If you sell a dev tool: what do most of your buyers pick, lifetime or monthly?
Someone left your team last month.
Can you say, right now, every repo they can still push to?
Most offboarding stops at "removed from the org". That's step one. The full check:
1. Remove them from the org (People page). They lose access to private forks of your private repos. Their local clones stay with them.
2. Open Outside collaborators. People added straight to a repo aren't org members, so step 1 misses them.
3. Org settings, Personal access tokens, Active tokens. Revoke the fine-grained tokens they created. Classic tokens don't show up here.
4. In each repo, open Deploy keys. Delete any key nobody can explain.
Step 4 is repo by repo. That's where old access hides.
I built a free tool for this: github-user-check. It runs locally, audits who has access to your GitHub repos, and removes access in bulk.
When someone leaves, who on your team removes their GitHub access, and how do they check it's done?
A Node.js vulnerability was reported yesterday. I wanted to know if our server was affected.
Instead of hunting through advisories and running commands by hand, I typed one thing into the @ctrlops_ai AI terminal:
"A Node.js vulnerability was reported yesterday. Search the web, find which versions are affected, check if our server is running one, and suggest a fix."
It searched the web, ran the checks on the server, and kept me in the loop the whole way. Running Claude Opus 5.5 under the hood.
Result: the server is vulnerable.
I said fix it. It patched Node, with my approval at every step.
Fix complete. Under 15 minutes. Full recording attached.
Did you check your server yet? π
π¨ Linux Infrastructure & SecOps Alert:
On 22 Oct, Trend Micro's Zero Day Initiative posted six Linux kernel security reports to public kernel mailing lists, before any fix was ready. Two of them are critical (CVSS 9.0) remote code execution bugs in ksmbd, the SMB file server built into the Linux kernel.
ZDI-CAN-33774: no authentication needed. If ksmbd is running and port 445 is reachable, an attacker can attack it remotely. This is the dangerous one.
ZDI-CAN-33549: in SMB Direct. It needs RDMA hardware and authenticated access to a share, so the risk is lower.
The other four reports cover skbuff, vmwgfx, nvmet and TCP.
Status: as of 22 Oct, neither ksmbd bug has a CVE number or a fix in the mainline kernel. Per The Hacker News, the bugs were found by TrendAI, Trend Micro's AI vulnerability scanner. RHEL does not ship ksmbd by default. NAS devices, homelabs and appliances that use the kernel SMB server are most at risk.
What to do today: Check whether it is running: lsmod | grep ksmbd and systemctl status ksmbd If it is there and you don't need it, stop it, unload the module and blacklist it. If you do need it, keep port 445 off the internet, allow only trusted networks, and apply the kernel update as soon as a fix ships.
Sources: Phoronix: https://t.co/iCWSZxwK05
The Hacker News: https://t.co/e9Z2aze7sx
I asked my AI assistant to send one message for me.
It couldn't. I had built it that way.
I'd just given it a tool for my messages. It can read and look things up. There's no send code in it at all, and its own rules say it never sends.
The first plan was different: it drafts, then sends after my yes. What I built is read-only. Sending stays with me.
An instruction like "never send" can be argued with or forgotten. A missing capability can't.
So if an agent must never do something, don't just tell it. Build the tool without that ability. Read access first, write access only when there's a reason.
The first person this rule stopped was me. I sent the message myself.
What's one thing your agent can read but should never be able to write?
Before you pick your next focus from a funnel table, check how it counts.
Five checks from our @ctrlops_ai funnel page in Mixpanel:
1. Count people, not page views. One person reading five pages is one visitor.
2. Take bots and AI agents out of visitors. Leave them in and every page looks worse at turning visitors into signups.
3. Credit each person to their first touch. And expect an "undefined" row: visits from before tracking started have no first touch.
4. When a row has only a handful of payers, read it as an early signal. Not a finding.
5. Note when you read it. Live data moves: the same table, read a few hours apart, gave slightly different totals.
Get these wrong and the table still looks clean. It just points you at the wrong thing.
Which of these does your dashboard get wrong today?