Choose an engine: eight AI systems I build.
Results: what this way of working has done.
Three numbers from my own jobs. Nothing rounded up, nothing borrowed from a vendor deck. The full story of each is on my work page.
Annual marketing and sales budget I've co-handled with the VP of Marketing, across JAIN Online, EdNxt.ai and about 20 colleges of the JAIN Trust, formerly USDC Global.
Where every rupee of wasted effort shows up.More leads from one organic push at JAIN Online than our usual Meta and Google volume.
One push, measured against paid.Each engine, taken apart.
The blueprint above shows the flow. This is what goes in, where a person signs off, what comes out, and the one rule that stops it going wrong. Pick an engine.
Machine steps in blue, the human checkpoints in amber. Tap a step for its blueprint.
Where humans stay in the loop.
"Human-directed" isn't a slogan. It's a design choice about how much the machine does before someone looks. Slide the dial and see what changes.
Amber is where I start every engine: a person sees every piece. Tap a level, or slide the dial.
What should you automate first?
Two questions sort almost every task: how often it comes back, and how much judgment it needs. Try one of yours, or tap an example.
Time-saved calculator: how many hours come back?
Put in your own numbers. The maths is plain and the assumptions are written underneath, so you can argue with them.
- A month is 4.33 weeks. A full-time month is 160 hours.
- The share is your guess. Start low, and raise it only after a pilot shows what the system can take.
- Review time is subtracted, because someone has to check the output.
- Tool costs and build time aren't included. They depend on your stack.
- Hours back aren't savings until you decide what those people do next.
Every tool, and what it's for.
The ones marked IN MY DAILY STACK are what I use every day. The rest I pick per job.
Tool map: which tool shows up where.
The same 38 tools, mapped against the eight engines. A blue dot means the tool does a job in that engine. The count on the right is how many engines lean on it, which tells you what to learn first. Amber dots mark tools in my daily stack.
Build it, buy it, or wire it?
Most teams don't need custom software. Some do. Answer three questions and the card that fits will light up.
The short version of the three cards: most teams land on wiring, with a person checking in the middle.
When the job is common and the tool already does it well.
- Scheduling posts, warming up email domains, transcribing calls
- You'd only be rebuilding what a product team maintains full time
- Watch for: one more login nobody owns
When the pieces exist but nothing connects them.
- Research in one tool, writing in another, publishing in a third
- n8n, Make or Zapier glue, plus AI steps and a human check in the middle
- Watch for: quiet breakages when an app changes. Someone must own it.
When the process is your edge, or your data can't leave.
- Your own voice rules, your own sources, your own approval steps
- Scripts and small apps written with Claude Code, living in your accounts
- Watch for: building before the process is stable
How AI systems fail, and the guard for each.
Every one of these is normal. None of them is a reason to stop. Each needs a specific guard built into the system, not a hope that someone notices.
Hallucination
Confident, fluent and wrong. A made-up stat, a source that doesn't exist, a feature your product doesn't have.
Brand drift
Week one sounds like you. Week ten sounds like everyone. Tone slides a little with every prompt tweak.
Slop
More output, less worth reading. Filler intros, lists of three, the same words every AI uses.
Prompt injection
Text hidden in a web page, email or document tells your AI to do something you never asked for.
Data leaks
Customer lists pasted into a personal chat account. Contracts uploaded to a tool nobody vetted.
Silent rot
An app updates, a login expires, a step fails, and the automation keeps "running" while doing nothing.
What people assume about AI at work, and what I see.
Four things I hear in almost every first call, and what the work has shown me instead.
Data and privacy, the India basics.
If your AI system touches customer names, phone numbers or emails, India's Digital Personal Data Protection Act, 2023 applies to you. Here's the shape of it, kept general.
- August 2023. Parliament passes the DPDP Act.
- November 2025. The government notifies the DPDP Rules. The Data Protection Board provisions start straight away.
- About a year later. Consent Manager provisions kick in.
- About 18 months later. Most duties on businesses apply: notices, consent, security safeguards, breach reporting and children's data.
- The stakes. Penalties go up to ₹250 crore for failing to take reasonable security safeguards.
This is a plain-English summary, not legal advice. Dates and duties come from the published Act, the 2025 Rules and news coverage of their notification. Check the current text on MeitY's site or with a lawyer before you rely on it.
What it means for an AI workflow
- Pasting personal data into an AI tool is processing it. You're still responsible for it, even when the tool is someone else's.
- You need a lawful reason, usually consent that is free, specific and informed, given after a clear notice of what you'll do with the data.
- People can ask to see, correct or erase their data, and must have a way to complain.
- Under-18s need verifiable consent from a parent or guardian.
- If there's a breach, the Board and the people affected must be told.
What I'd put in place by default
- Business plans with data controls for anything with customer data in it.
- Mask or drop names when the task doesn't need them.
- Only message people who signed up for it.
- Keep a one-page list of which tool sees which data.
The checklist before anything ships.
Run new AI output through this every time until it passes ten times in a row. After that, the person checking can move to spot checks. Tick as you go; your ticks stay in this browser.
A few prompt patterns worth stealing.
Not magic words. Structures that make output checkable. Copy one, swap the brackets for your details, and see the difference. More of these go into the Applied AI notes.
How this site's blog engine works.
The Field notes on this site run on the same idea I sell: AI does the heavy lifting, I make the calls. Here's the actual loop.
Topics in
I drop a line per post into an inbox: the world, a working title, my notes and where my files and photos are.
Drafted every morning
Claude writes each new topic as a draft, never published, following a playbook: answer first, real questions as headings, sources at the end.
No guessing
If a fact can't be verified, the draft carries a [CHECK] note instead. Moto news needs two independent sources, the maker's or the government's first.
I approve
I read the draft, edit what I like, and mark it approved. Or I hold it.
Published on schedule
The next approved post goes live on its slot, one per run, and then gets checked on the live site.
Applied AI notes: how I actually use the tools.
Rollout plan: the first ninety days, month by month.
One engine, done properly, beats five half-built ones. This is the order I'd follow with any team.
First month
- Sit with the team and list every task that repeats weekly
- Sort them with the frequency and judgment test above
- Write the voice guide and the approval rules
- Build the first engine and run it alongside the old way
Second month
- Fix what broke in the first month, add logs and counts
- Move reviewers from full checks to spot checks where it has earned it
- Start the second engine, usually the one that feeds the first
- Measure hours against the baseline
Third month
- Document every engine so the team can run it without me
- Train the owners, not just the users
- Set a monthly review for drift, cost and results
- Decide what the freed-up hours go to
Glossary: the words, without the jargon.
Every term on this page, in one line each. Search if you're looking for one.
Questions buyers ask me.
What does "human-directed, AI-built" actually mean?
AI does the building: research, drafts, edits, versions, scheduling. A person sets the direction and signs off at fixed checkpoints, like facts, money, anything public and sensitive replies. The machine never has the last word on those.
Will AI-written content hurt my Google rankings?
Google's stated position is that it rewards helpful content however it's made, and acts against content made mainly to game rankings. So the risk isn't AI, it's thin, unchecked content. The content engine is built around sources, real search queries and a human check for that reason.
How much time can a team realistically save?
It depends on how much of the week is repetitive. My one measured result is a 60% shorter production cycle at Gnani.ai, and that's a result, not a promise. Use the hours calculator with a conservative share to get your own number.
Do we need to buy new tools?
Usually not many. Most engines run on tools teams already pay for, plus one or two AI tools and some glue. The tool matrix shows which tools carry the most weight.
Is it safe to put customer data into AI tools?
Only with care. Use business plans with data controls, share the minimum, and follow India's DPDP Act on consent and security. The privacy basics above cover the shape of it. Get legal advice for anything specific.
What if the AI gets something wrong?
It will, sometimes. That's why every engine has a checkpoint before anything goes out, and why unverified facts get flagged instead of guessed. The failure modes section lists the guard for each kind of mistake.
How do we start working together?
Tell me the outcome you want. You can scope it and see the cost on the hire page, or read what I've done before first.
Work with me: three ways to start.
More from the Saptaverse.
Applied AI is one corner of the site. Marketing after AI lives in Bazaar. The rest is machines, travel, India, the mind, design and food.