Indrani Som

[Indrani Som @indrani_som2004]

I spent the last week building an AI chatbot SaaS from scratch.

🤖 Crawl any website
🧠 Generate embeddings
🔍 Semantic search with Qdrant
💬 RAG-powered answers
⚡ Deploy with one script tag
Stack:
Next.js • Node.js • Playwright • BullMQ • Redis • Qdrant • NVIDIA LLM
This has been one of my biggest AI engineering projects yet.
🧵 Here's how it works...

[Indrani Som @indrani_som2004]

@Suman_N_Jain Thankuuuu 👀

[Indrani Som @indrani_som2004]

I spent the last week building an AI chatbot SaaS from scratch.

🤖 Crawl any website
🧠 Generate embeddings
🔍 Semantic search with Qdrant
💬 RAG-powered answers
⚡ Deploy with one script tag
Stack:
Next.js • Node.js • Playwright • BullMQ • Redis • Qdrant • NVIDIA LLM
This has been one of my biggest AI engineering projects yet.
🧵 Here's how it works...

[Indrani Som @indrani_som2004]

@Anjishnu46 @hanqing1120 Haha 😂on my way ….

[Indrani Som @indrani_som2004]

@jeffreyon_ These appreciations will be the motivation for my next project 🙂‍↔️

[Indrani Som @indrani_som2004]

@hanqing1120 😂 Challenge accepted! Looks like “One-click Boss Weekly Report Generator” just got added to the roadmap for my next post.

[Indrani Som @indrani_som2004]

Thanks! The focus was definitely making deployment effortless. For chunking, I first clean the extracted content and remove duplicate sections like repeated headers, footers, and navigation that appear across multiple pages. That keeps the chunks cleaner and helps the retriever focus on the actual content. I’m still refining it as I test against more complex websites.

[Indrani Som @indrani_som2004]

@themanasai Thanks! That’s a great suggestion. Hybrid search is definitely something I want to explore as Scrappy scales, especially for exact feature and product name retrieval.

[Indrani Som @indrani_som2004]

@Anjishnu46 Here is the repo link : https://t.co/KNHYUfpfpI

[Indrani Som @indrani_som2004]

I noticed something while building websites.

Almost every client asks:

“Can we add an AI chatbot?”

Mostly either we make it using n8n or any custom third party site where we need to update a lot of information about our site so what if

We just add a simple script tag and the bot itself crawls to your entire site, makes its own knowledge base and answers user’s query accordingly ?

Will you use a chatbot like that in your website ? Tell me in the comments 👀

[Indrani Som @indrani_som2004]

Atleast with job we can cry with a cold coffee :)

[Indrani Som @indrani_som2004]

Day 3 of shitposting about my journey to learning new things

Started building Scrappy AI 🚀

The goal is simple: add one script to any website and get an AI chatbot that can answer questions using the website's own content.

One thing I learned today:

Scraping ≠ Crawling

Scraping extracts data from a single page.
Crawling discovers and visits all the internal pages of a website before extracting their content.

For an AI chatbot to truly understand a website, crawling is the right approach. Now onto building the crawler, followed by chunking, embeddings, Qdrant, and RAG.

#AI #RAG #Qdrant #BuildInPublic

[Anjishnu Ganguly @Anjishnu46]

Hey , @rapidoapp_ , I am frequently encountering an issue where drivers call prior to arrival demanding fare amounts significantly higher than the app's estimate. When I decline, they refuse to pick me up and insist that I cancel the trip.

One driver almost asked for a double fare.

[Indrani Som @indrani_som2004]

@xPraveen07 Late night = peak hours for writing and making shit

[Indrani Som @indrani_som2004]

New Full stack becoming
React + Fast API + Qdrant + ‘Powered by AI’ ? 😂

[Ojas Sharma @OjasSharma276]

First it was the MERN stack. Now it's Data Science and AI/ML

[Palak🎀 @Palak3312]

I don't want a luxury car. I want a coding screen like this. 👨‍💻

[Indrani Som @indrani_som2004]

Hey @hoichoitv

Noticed a small UX issue while using the platform.

When a request fails, the UI only shows “Request failed with status code 400”, which isn’t very helpful for users. The backend likely already returns a meaningful error message, but it isn’t being displayed.

Showing the actual message from the API (instead of the generic Axios error) would make debugging and the overall user experience much better.

Just a small suggestion. 😊

[Indrani Som @indrani_som2004]

@Anjishnu46 @thsottiaux A much needed feature fs .

[Indrani Som @indrani_som2004]

Day2 of learning new things and posting about it

So Today let's learn about RAG in few words

What's RAG?

LLMs answer from memory.

RAG says, "Wait... let me check the docs first." 📚🤖

Retrieve the right information ➜ Generate the answer and stop the hallucination

Basically, AI with receipts. 😄

#AI #RAG #LLMs #AIEngineering

[Indrani Som @indrani_som2004]

@Anjishnu46 Thank you 😇

[Indrani Som @indrani_som2004]

I am trying to start my journey towards learning AI/ML since a long time now but each time I start I just lost my motivation so I decided to post small ideas or parts of my learning through visuals and some examples each day I learn something new . So this is Day1 of me posting about a small idea

How "Frequently Bought Together" Actually Works

Imagine you open your mobile to order a bread from Blinkit and buy a loaf of Bread.
Immediately, Blinkit suggests Butter with a sign Frequently Bought Together

How did it know?
Step 1: Unsupervised Learning
There are no labels or correct answers.

The system simply looks at millions of past shopping carts to discover hidden patterns.

Step 2: Association Rule Learning
It asks one simple question:
"What products are usually bought together?"
Example:
Customer 1 → Bread + Butter
Customer 2 → Bread + Butter + Milk
Customer 3 → Bread + Eggs
Customer 4 → Bread + Butter + Jam

Step 3: Apriori Algorithm
Apriori scans all transactions.
It counts how often different products appear together.
It discovers:
Bread + Butter appears very frequently.
So it creates the rule:
Bread → Butter

Step 4: FP-Growth Algorithm

Apriori works well, but scanning millions of transactions repeatedly becomes slow.

FP-Growth stores the data in a compact tree structure.
Instead of checking everything again and again, it finds the same patterns much faster.

Final Result
When a new customer adds Bread to the cart...
The recommendation system already knows:
People who buy Bread often buy Butter.
Frequently Bought Together
Ans:🧈 Butter

This diagram is made using https://t.co/Jp4e5omPMB