A plain-language guide to our three in-browser demos: what each does, how to try it step by step, and what to look for.
Before you start
Three small demos, no sign-up.
Each demo shows one idea we build into AI products, in about two minutes. This page tells you what to click, what you should see, and what each demo cannot do.
They run only in your browser. Nothing is sent or stored.
The data is made up, and no real AI model is called.
Before text is sent to an AI model, personal details such as emails, phone numbers, Aadhaar, PAN and card numbers are swapped for placeholders like [EMAIL_1]. The real values are kept aside in a "vault" and put back into the reply afterwards. Every step is logged without the original values.
Try it in about two minutes
Press "Use sample text". The sample text and a name are filled in and masked for you.
Read step 2, "What the model would see". Each sensitive value is now a coloured placeholder. The email appears twice in the sample and gets the same placeholder both times.
Open "The vault" under it to see what each placeholder stands for. Tick "Hide original values" to blur them.
In step 3, press "Restore values". The real details return in the reply.
Now add a made-up placeholder to the reply, such as [EMAIL_9], and press "Restore values" again. It is flagged as not in the vault and left as it is.
In step 4, press "Download audit log (JSON)" and open the file. It lists types and counts only, never the values.
What to look for
The same value always gets the same placeholder, so a model can still tell who is who.
Names are only hidden when you list them in "Names to hide". Names have no fixed pattern.
Try your own text, but use made-up details only. Nothing is sent anywhere, but it is a good habit.
What it does not do
It matches formats, not meaning. A card number is only caught when it passes the card checksum, and the vault lives in the page, so it is gone when you refresh.
An AI can only read so much at once. This demo plays the packer that decides which tools, earlier messages and help articles fit into that limited space for a customer question, and leaves the rest out.
Try it in about two minutes
Under "The customer asks", pick "Charged twice". The question appears in the box.
Under "Size of the suitcase", drag the slider all the way to Tiny. Read the sentence below it: very little is packed.
Drag the slider up to Roomy. More items come in, but unrelated ones stay out even when there is plenty of space.
Click "Change address". A different tool and a different help article are packed, because the question changed.
Look at "How the room is used". The bar shows how the space is shared, including a slice kept free for the answer.
Scroll to the three cards. Ticked items are packed, crossed items are left out, with the reason. Open "See the exact text the AI would receive" at the bottom.
What to look for
Items that switch between packed and left out flash briefly as you move the slider.
Tools and articles that are unrelated to the question are dropped, even when there is room.
Sizes are in tokens, which are small chunks of text, roughly three quarters of a word each. Here they are estimated.
What it does not do
Relevance is simple word matching. A real system would use embeddings and the model's own tokenizer, and would summarise old messages instead of dropping them.
Not every question needs the expensive AI. A router reads each question and sends easy ones down a fast, cheap lane, and hard ones to a slower reasoning model. The demo shows the clues it notices and the cost and risk of different settings.
Try it in about two minutes
Click the example "Opening hours". It goes to the fast lane, with a "common look-up" clue.
Click "Compare plans". It goes to the thinking lane, because it is long and asks for reasoning.
Read "What the router noticed". Each clue has points, and the points add up to the difficulty score.
Click "Legal question". With the dial at 2 points it goes to the thinking lane. Now move the dial in step 2 to 3. The same question drops to the fast lane, which is risky.
In step 3, watch the cards as you move the dial from 2 to 5. Savings go up, and so does the "risky" count.
Click any question in the list to load it into step 1.
What to look for
With the sample questions at 2 points, the router is about 48% cheaper than always using the big model, with no risky calls. At 3 points it is about 76% cheaper, but 4 hard questions slip into the fast lane.
Cheaper is not free. The more you save, the more mistakes get through.
Costs and times are illustrative numbers, not real prices or measurements.
What it does not do
The clues are simple word rules. A real router often uses a trained classifier, and retries on the reasoning model when the fast model is unsure.
A small chunk of text that an AI model reads. Roughly three quarters of a word.
Context
Everything an AI is given to read before it answers: the question, instructions, tools, earlier chat and documents.
PII
Personally identifiable information: details that can point to a real person, such as an email, phone number or ID number.
Placeholder
A stand-in such as [EMAIL_1] that replaces a real value until it is put back.
Vault
The private store that remembers which placeholder stands for which real value.
Audit log
A record of what happened and when. Here it keeps types and counts, never the values.
Router
A small piece of logic that decides which AI model should handle a request.
Reasoning model
A slower, more expensive AI model that is better at hard, multi-step questions.
Relevance
How closely an item matches the question. In these demos it is a simple word-matching score.
Questions
Good to know.
Is anything I type sent to you or to an AI?
No. The demos run entirely in your browser. They make no network calls and store nothing, so refreshing the page clears everything.
Is a real AI model running?
No. Each demo is a rule-based simulation on made-up data, so it is free to use and cannot be abused. It shows the idea, not a production system.
Why are some results a bit odd?
The rules are deliberately simple, so you can see how the decision is made. Real systems use trained models and are more accurate, and the pages list what they leave out.
Do I need an account or a special browser?
No. Any recent browser on a phone or computer works. Use the sun or moon button in the header to switch between light and dark.
Can I use the keyboard?
Yes. Use Tab to move between controls, Space or Enter to press buttons, and the arrow keys to move a slider.
Can you build something like this for my product?
Yes. These are small versions of ideas we build into real products. Get in touch and tell us what you have in mind.
Have an idea? Let's build it properly.
Tell us what you want to build or automate. We'll reply with practical next steps.
