About Todkett: an automation lab for people learning AI automation
Todkett is an automation lab for people learning AI automation: founders, operators, freelancers and students who want to automate their own work in n8n. We publish workflows you can copy step by step, try to break each one before it goes up, and show the checks it passed. Every AI step is marked.
30 minutes about the workflow you are building. A written summary with the next step within 1 working day.
01Who answers your call
You talk to the person who builds and tests these workflows.
The person on your call is the person who designs, builds and tries to break every workflow on this site.
Nitin Sarkar
- Role
- Founder, builds and tests every workflow
- Based in
- Bangalore
- On your call
- The same person: nobody else answers it
02How we build
How we build, and how you can too.
Rules around every AI step. AI reads, judges or drafts. Fixed rules check, decide and route. A person approves anything that goes out. Every workflow on this site follows this shape, and the one you build can too.
The shape every workflow takes
The pattern, not a sample run. Each workflow page draws its own line.
The IDs on the line mark where a check applies. Each one is explained on the test bench below.
Every step is labeled
On every workflow page each step is marked as an AI step, a rule step or a person step, so you can see where a model is used and where it is not. Label your own steps the same way.
AI stepRule stepPerson step
Rules before the model, code after it
Fixed rules filter, cap and dedupe what reaches the AI step, and code checks the answer it returns. If the answer is broken, or the model is down, the item falls back to a safe result: needs review, a template answer or a default value.
A person approves what goes out
Nothing reaches a customer, a candidate or a money system without a person, unless a fixed rule allows it, for example a discount inside a limit you set.
Outside text is data, not instructions
Text a workflow reads, such as an email, a resume or an invoice, is passed to the model as quoted data. Every AI step that reads outside text is tested with instructions hidden in the input.
What a workflow touches, and where it is kept
Each workflow page names every system it reads from and writes to, and where each result is kept. Run your copy on sample data before it touches your real data. We show what we tested, never the test data.
Limits and a failure alert
Each workflow has limits on input size, items per run, sending speed and AI spend. If a run fails, a named person gets an email saying which step failed. In your own build, that person is you.
Model changes are tested again
Each AI step uses a model chosen for that step. If we change the model or the prompt, the tests run again before the change goes live. Do the same when you swap a model in your copy.
Built in n8n
We design and build in n8n, and every workflow here is drawn so you can rebuild it there. It connects to the tools you already use, such as email, a CRM and spreadsheets.
03Stress test
Before you trust a workflow, try to break it.
Every workflow on this site faces the same 8 checks, and its page shows which ones it passed and when. They are the 8 checks we run, and you can run too, on anything you build.
A check marked Not needed says why. If we change the AI model or the prompt, the tests run again before the change goes live; do the same in your own build. We show what we tested, never the test data, except the made-up example on each page.
The test bench
The standard, not a result. Every lamp here is unlit: each use-case page lights these signals with its own recorded run.
- Passed
- a recorded test run passed
- Not yet
- still open
- Not needed
- with one of the reasons named
- ST1: link to this check
Test set
We run a fixed set of labeled cases, with the right answer written down first, and compare every result.
- On its page, a pass reads
- a count labeled cases, a count runs in a row
- Not needed only for
- No reason. Every workflow faces it.
- ST2: link to this check
Prompt attack
At least one test hides instructions in the input. The workflow must ignore them or flag them.
- On its page, a pass reads
- a count attack cases: instructions hidden in the input
- Not needed only for
- “No AI step”“AI sees only numbers our code computed”
- ST3: link to this check
Run twice
Sending the same input twice, or rerunning after a crash, never doubles a record, a message or a payment.
- On its page, a pass reads
- Same input sent twice: one record
- Not needed only for
- “Returns a result only: writes and sends nothing”Free tools: “Runs in your browser: no server”
- ST4: link to this check
Failure recovery
A failed call is retried a set number of times. If it still fails, the item is parked for a person. Nothing is dropped silently.
- On its page, a pass reads
- Failed calls retried a count times, then parked for a person
- Not needed only for
- No reason. Every workflow faces it.Free tools: “Runs in your browser: no server”
- ST5: link to this check
Load and limits
We run it at its own maximum batch size, and it has limits on input size, items per run, sending speed and AI spend.
- On its page, a pass reads
- Full batch of a count items tested; limits on size, batch and AI spend
- Not needed only for
- No reason. Every workflow faces it.
- ST6: link to this check
Failure alert
If a run fails, a named person gets an email saying which step failed.
- On its page, a pass reads
- Failure email names the step that failed
- Not needed only for
- No reason. Every workflow faces it.Free tools: “Runs in your browser: no server”
- ST7: link to this check
Human sign-off
Nothing reaches a customer, a candidate or a money system without a person, unless a fixed rule allows it.
- On its page, a pass reads
- Drafts onlyor Risky items go to a personor Fixed approval rule: the rule in plain words
- Not needed only for
- “Returns a result only: writes and sends nothing”“Stays with you”Free tools: “Runs in your browser: no server”
- ST8: link to this check
Rules around the AI
Fixed rules run before the AI and check its answer after. If the AI fails or is down, the workflow falls back to a safe result.
- On its page, a pass reads
- Rules before the AI, code checks after, safe fallback
- Not needed only for
- “No AI step”
No upload
During every test run, no file or text left the browser.
- On its page, a pass reads
- No file or text left the browser in a count test runs.
- 8 checks for every workflow
- 4 always apply
- 4 may read Not needed, with the reason shown
- 1 more for free tools
Not needed, and why
A check may be skipped for one of these five reasons only, and the page names the reason beside the check.
| Reason, as the page shows it | May mark |
|---|---|
| No AI step | ST2 Prompt attackST8 Rules around the AI |
| AI sees only numbers our code computed | ST2 Prompt attack |
| Returns a result only: writes and sends nothing | ST3 Run twiceST7 Human sign-off |
| Stays with you | ST7 Human sign-off |
| Runs in your browser: no serverFree tools only | ST3 Run twiceST4 Failure recoveryST6 Failure alertST7 Human sign-off |
The test ledger
One row for each live use case and free tool, read from its recorded test runs. No number here is typed by hand, rounded or estimated.
| Page | Checks passed | Labeled cases | Prompt-attack cases | Runs in a row | Last run |
|---|---|---|---|---|---|
| AI resume screening | 0 of 8 | None | None | None | None |
| Inbound lead qualification | 0 of 6 | None | None | None | None |
| Invoice data extraction | 0 of 8 | None | None | None | None |
| Support ticket triage | 0 of 7 | None | None | None | None |
| Weekly KPI report | 0 of 6 | None | None | None | None |
| Social media post drafts | 0 of 8 | None | None | None | None |
| Call notes to CRM | 0 of 7 | None | None | None | None |
| Proposal generator | 0 of 8 | None | None | None | None |
| Competitor research deck | 0 of 7 | None | None | None | None |
| Customer health score | 0 of 7 | None | None | None | None |
| Hiring signal lead finder | 0 of 8 | None | None | None | None |
| AI video auto-posting | 0 of 8 | None | None | None | None |
| Automation ROI calculator | 0 of 4 | None | None | None | None |
| ATS resume checker | 0 of 4 | None | None | None | None |
| PDF toolkit | 0 of 4 | None | None | None | None |
| Total | 0 | 0 | 0 |
04Where we work
Based in Bangalore, working with learners worldwide.
Todkett works from Bangalore (Bengaluru), in India Standard Time, with learners anywhere, remotely.
Some places have their own page, with the local rules that change a workflow and the hours that overlap with ours: Bangalore, Dubai, Singapore and London. All of them are under where we work.
Delivered projects are described by industry, without client names or figures.
Call hours, and how they line up with your day, are on the booking page.