Nitij Taneja

AI × Neuroscience Research | Agentic & Multi-Agent AI | LLM Engineering GraphRAG Scalable AI | Intelligent & Cognitive Systems | Integrated @ VIT

[Nitij Taneja @TNitij]

know I just posted one 😅 but maintaining the timeline!

Another paper accepted at IEEE COMPUTINGCON 2026 🎉
Presenting on Sep 7–8, likely online due to work constraints.

3 in a row this year. More to go! 🚀📚

#Research #IEEE #AI

[Nitij Taneja @TNitij]

A little late, but worth sharing! 🚀

Presented my paper “A Multi-Agent Graph RAG Framework with React and A2A Protocol” at IEEE INSTcon 2026, NIT Rourkela.

Exploring Multi-Agent Systems, Graph RAG & A2A communication. 🤖

#AI #GraphRAG #AgenticAI #IEEE

[Nitij Taneja @TNitij]

Reading this fascinating 2024 paper on LLMs and the human brain 🧠🤖
The interesting part: better-performing LLMs show more brain-like language processing, reach similar neural representations in fewer layers, and seem to benefit strongly from context.
@Steph_Bickel

[Nitij Taneja @TNitij]

Just went through “Dense Passage Retrieval for Open-Domain Question Answering” — loved how a simple dual-encoder setup, trained with in-batch negatives, can beat traditional BM25 so convincingly. A great paper on making dense retrieval actually work in practice.

[Nitij Taneja @TNitij]

@HarperSCarroll @huggingface @mervenoyann Iron Man built his first suit in secret and kept it off the grid. Open models are basically that energy for AI.

[Nitij Taneja @TNitij]

@finkd Open weights is always a win for the open-source community. Excited to see what people build with Muse Glimmer and Muse Spark 1.2. 🚀

[Nitij Taneja @TNitij]

Ever wondered how apps like Instagram load profile pictures so fast for millions of users?

In this video, I explain the complete pipeline:
📤 S3 Uploads
⚙️ Image Processing
🌍 CloudFront CDN
🏷️ URL Versioning
✨ BlurHash

System Design made simple.

#AWS #CloudFront #S3 #System

[Nitij Taneja @TNitij]

Most people think HuggingGPT is another AI model.

It isn't.

It's an AI manager that plans tasks, picks the best expert models from Hugging Face, combines their outputs, and returns one final answer.

#AI #HuggingGPT #LLM #AIAgents

[Nitij Taneja @TNitij]

@ClementDelangue As a student researcher, I couldn't agree more. Open datasets and open-weight models have shaped much of my learning. Open ecosystems don't just build better AI—they build better researchers.

[Nitij Taneja @TNitij]

@ylecun @ClementDelangue Interesting perspective. The recent success of open-weight models suggests that openness is still one of the strongest catalysts for research progress, even as deployment becomes more controlled.

[Nitij Taneja @TNitij]

Our research paper is now published! 🎉

A Spatio-Temporal Approach to Sales Forecasting

DOI: https://t.co/qSJ5FbkT3T

#AI #DataScience #Research

[Nitij Taneja @TNitij]

Most people think 1 token = 1 word.
In this video I explain:
• What AI tokens actually are
• Tokenization
• Token IDs
• Image & Audio Tokens
• Input vs Output Tokens
• Context Window
• API Pricing
If you use ChatGPT or build AI apps, you'll finally understand how LLMs work

[Nitij Taneja @TNitij]

Hybrid Database Architecture 🗄️

Modern apps don't store images inside PostgreSQL.

✅ PostgreSQL → Users, Posts, Metadata
✅ AWS S3 → Images, PDFs, Videos

Here's a simple workflow 👇

#AWS #PostgreSQL #SystemDesign #Backend #WebDevelopment

[Nitij Taneja @TNitij]

What happens when AI agents face hundreds of tasks at once?
Today's agents struggle with concurrency—maintaining context, prioritizing, and preserving state under pressure.
Next-gen systems need better coordination, memory, and real-time decision making.
@hwchase17 @swyx
#ai

[Nitij Taneja @TNitij]

What if AI agents didn't have to re-reason through every routine task?

A procedural memory layer could transform past reasoning into reusable execution patterns—making agents faster, more efficient, and less computationally wasteful.
@fchollet @hwchase17 Thoughts 👀
#ai #agent

[Nitij Taneja @TNitij]

What if the next in AI isn’t bigger models, but smarter attention?
The prefrontal cortex filters information before it overloads the brain. Maybe LLMs need the same biologically inspired resource management instead of endless scaling.
@GaryMarcus @bengoertzel Thoughts? 👀
#ai

[Nitij Taneja @TNitij]

GloVe proves meaning is in the ratio: Ice/Steam co-occurrence ratios perfectly capture thermodynamics phases.

Is language just a statistical matrix? #nlp #Glove #Ai #stats

[Nitij Taneja @TNitij]

📄 "Distributed Representations of Words and Phrases and their Compositionality" (2013) extended Word2Vec with phrase learning. 🚀

But if it still couldn't truly represent entire sentences or long documents, did Transformers finally solve it—or are there still gaps? 🤔

[Nitij Taneja @TNitij]

Can we encode thoughts as vectors & math?
Word2Vec (Mikolov et al. 2013) made king - man + woman ≈ queen real by turning words into distributed vectors.Simple log-linear model, Huffman softmax, CBOW & Skip-gram.Can we do the same for full thoughts? Or already?#Word2Vec