Tanmay R Thombare

Engineer. Builder. Explorer

Building in public

[Tanmay R Thombare @Tan_builder]

The evolution of web visibility:
​2015: "How do I rank on page 1?" (SEO)

2020: "How do I get the featured snippet?" (AEO)

2026: "Why didn't Chatgpt cite my benchmark?" (GEO)
​
To businesses and product builders out there how important are AEO and GEO??

[Tanmay R Thombare @Tan_builder]

@darioguariglia I have been building ai agents too. Now exploring Agentic ai security and building scalable agents

[Tanmay R Thombare @Tan_builder]

Hey Founders, Engineers and Tech enthusiasts !!
I’m Tan. I’m an engineer who loves building things.

If you’re into AI and love building things, let’s connect.
What are you working on? 🤝

[Tanmay R Thombare @Tan_builder]

@darioguariglia Hey dario!! Great to connect with someone doing same work. Curious, what are you building??

[Tanmay R Thombare @Tan_builder]

@PaulShipsAI Done man. Companion app? Interesting... would love to know more about it.

[Tanmay R Thombare @Tan_builder]

@GShip223 Will try that soon.
True, after a certain point GPU is a must.

[Tanmay R Thombare @Tan_builder]

LLM inference is HARD!! atleast for me.

Been learning about LLM inference and as i digged deeper i understood its a rabbit hole. So many concepts so many tradeoffs. As i am going further
I want to learn it hands on,
Is self hosting a model on Raspberry Pi a good idea???

[Tanmay R Thombare @Tan_builder]

@entretokens 1B at Q4 in llama.cpp sounds a good start.
Seeing why decode is memory bound, these are the exact reasons i want to try these things myself.
Helps to get the intuition!

[Tanmay R Thombare @Tan_builder]

@CodeByPoonam Looks like AI is going to replace pet's too...

[Tanmay R Thombare @Tan_builder]

Here is the starting point. Built the basic app. Will further improve it based on the data provided by business (client).
Will soon start to implement features, improved designs, optimizations etc.

[Tanmay R Thombare @Tan_builder]

Did a small experiment.
Posted same content.
1. Written by me. Purposely did some mistakes.
2. Generated by AI. 100% correct grammer.

The post that was written without using AI shows more impressions.

Curious if one should use AI for writing??

[Tanmay R Thombare @Tan_builder]

As open source models are getting better and better I've realised how impotant self hosted models will be for businesses to keep data private hence. Today I am diving into LLM inference today.
Particularly for case of self hosted open weight models.
Will share my learnings soon

[Tanmay R Thombare @Tan_builder]

As open-weight models get better, I think self-hosted inference will become increasingly important for businesses that need control over their data.
Today, I'm diving into LLM inference, specifically serving open-weight models on self-hosted infrastructure.

[Tanmay R Thombare @Tan_builder]

Building a custom inventory management system for another business. (Fully local)
Tauri + sqlite.
I will be sharing how this progress✌️

[Tanmay R Thombare @Tan_builder]

@AnshikaK7 I know what a python GIL is.
But most important...
I know what is JSON😂😂

[Tanmay R Thombare @Tan_builder]

An LLM can't actually browse the web or execute code by itself.
It generates a response requesting a tool call (just generates text). Your code reads that response, runs the tool, and sends the result back to the LLM.

The model provides the decision. Your code does the execution

[Tanmay R Thombare @Tan_builder]

Pure speed hack!!!

Just learned how KV caching speeds up LLM inference and it’s genius.

​Instead of recomputing attention for every past token at each step, it stores the Keys & Values in memory.

​You trade VRAM to drop latency from quadratic to linear.

[Tanmay R Thombare @Tan_builder]

AI agents like Google’s Big Sleep and CodeMender aren’t just flagging code flaws. They can uncover zero-day vulnerabilities and automatically draft patches before attackers even spot them.
​Proactive AI defense is officially here.
​Read more: https://t.co/UWDnfBifeB

[Tanmay R Thombare @Tan_builder]

Google...always google!!!!
Google's research side is just too good.

[Superman @thesupermannx]

Researchers mathematically proved Transformers are computationally universal.

They applied the concept of Kolmogorov complexity, the mathematical absolute limit for perfect data compression and algorithmic generalization.

And it proved that Transformer encoders are computationally universal.

In plain English: they proved that there is a mathematical state where an AI model becomes perfectly compressed and generalizes flawlessly to new data.

The architecture is theoretically flawless.

But there is a catch. And it changes how we look at every AI lab on earth.

When the researchers built a specific objective to force the AI to find this mathematically optimal, perfectly generalized state... the Transformer was entirely capable of holding it.

But the optimizer couldn't find it.

Standard training methods, the exact same optimizers used to build ChatGPT, Claude, and Gemini, completely failed to discover this perfect solution from a random start.

We are staring at a massive, structural disconnect.

Transformer architecture is mathematically capable of universal, optimal intelligence. The engine is perfect.

But the steering wheel is broken.

Our training methods are simply too primitive to unlock what the model is actually capable of.

We don’t need to invent a new architecture to reach the next leap in AI reasoning.

Intelligence is already buried inside the models we have right now.

[Tanmay R Thombare @Tan_builder]

3. Tool Shadowing: A rogue MCP server mimics descriptions to get picked over trusted tools.

4. Indirect Injection: Data fetched by a tool carries hidden commands that steer the model.

5. Excessive Permissions: Broad access turns minor injections into major data leaks.

[Tanmay R Thombare @Tan_builder]

Spent some time analyzing MCP and MCP security.
Giving LLMs access to local and remote tools and APIs is powerful, but it opens massive risks.

​Here are 5 critical MCP attack vectors you need to know: 👇

[Tanmay R Thombare @Tan_builder]

1. Tool Poisoning: Hidden prompt injections in tool descriptions, this is most straight forward one.

2. Rug Pulls: A tool alters behavior on the remote server after you approve it.
Similar to tool poisoning but context/prompt changes after approval.